From Turing's question to transformers, LLMs, agents, and the collapsing cost of intelligence, 1943–2025. Click any paper to expand.
AI Foundations & LLMsEvals & BenchmarksML & EconometricsText as DataPrediction, Fraud & DecisionsAI & Asset PricingEconomics of AIGenAI at Work
Analytical / Theory & Methods
Empirical / Evidence & Applications
1940s
AnalyticalOrigins
A Logical Calculus of the Ideas Immanent in Nervous Activity
McCulloch & Pitts
Bulletin of Mathematical Biophysics, 1943
The FIRST mathematical model of a neuron. Showed that networks of simple on/off units can compute any logical function. Before computers even existed, this paper established the idea that thinking might be computation — and that networks of simple units might do it.
The intellectual ancestor of every neural network, from the perceptron to GPT.
1943
AnalyticalOrigins
A Mathematical Theory of Communication
Shannon
Bell System Technical Journal, 1948
Invented information theory in one paper: bits, entropy, channel capacity, redundancy. Defined information as reduction of uncertainty and made it MEASURABLE. Language models are, at their core, machines trained to minimize Shannon's cross-entropy on text.
Every measure of information content — in markets, in disclosures, in language models — descends from this paper.
1948
1950s
AnalyticalOrigins
Computing Machinery and Intelligence
Turing
Mind, 1950
Asked the question that started the field: “Can machines think?” — and replaced it with an operational test: can a machine's conversation be distinguished from a human's? The Turing Test reframed intelligence as behavior rather than essence, and anticipated machine learning: “Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's?”
Seventy-five years later, LLMs passing casual Turing Tests made the paper's question urgently practical.
1950
AnalyticalOrigins
A Behavioral Model of Rational Choice
Simon
Quarterly Journal of Economics, 1955
Introduced BOUNDED RATIONALITY: real decision-makers do not optimize, they SATISFICE — search until an option is good enough, then stop. Simon (a Turing Award winner AND an economics Nobel laureate) is the human bridge between economics and artificial intelligence: he spent the rest of his career building AI programs that mimicked human problem solving.
The reason AI matters for economics: if humans are computationally constrained, machines that relax the constraint change behavior, markets, and organizations.
1955
AnalyticalNeural Networks
The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain
Rosenblatt
Psychological Review, 1958
The first LEARNING machine: a neural network that adjusts its own weights from examples. Rosenblatt built it in hardware — the Mark I Perceptron — and the New York Times promised machines that would “walk, talk, see, write, reproduce itself and be conscious of its existence.” The hype cycle is not new.
Every gradient update in every deep network is a descendant of the perceptron learning rule.
1958
1960s
AnalyticalNeural Networks
Perceptrons
Minsky & Papert
MIT Press, 1969
The book that (almost) killed neural networks. Proved mathematically that single-layer perceptrons cannot represent simple functions like XOR — and expressed skepticism that multi-layer versions could be trained. Funding collapsed; the first AI WINTER followed.
A cautionary tale about proving limits of the wrong model class. The multi-layer networks they doubted now write poetry and pass bar exams.
1969
1980s
AnalyticalNeural Networks
Learning Representations by Back-Propagating Errors
Rumelhart, Hinton & Williams
Nature, 1986
The paper that revived neural networks. Backpropagation — applying the chain rule to send error signals backward through layers — made training MULTI-layer networks practical, answering Minsky & Papert's challenge. Hidden layers learn internal representations that no one programmed.
Still the training algorithm of essentially every deep network, including the LLMs. Hinton received the Turing Award (2018) and the Nobel Prize in Physics (2024) for this line of work.
1986
2000s
AnalyticalStatistical ML
Random Forests
Breiman
Machine Learning, 2001
Average hundreds of decorrelated decision trees and variance collapses while bias stays low. Random forests were embarrassingly simple, nearly tuning-free, and for two decades the strongest off-the-shelf learner on tabular data.
Still the first serious benchmark in most applied prediction papers in accounting and finance — and frequently still the winner on structured financial data.
AnalyticalStatistical ML
Statistical Modeling: The Two Cultures
Breiman
Statistical Science, 2001
The manifesto. Breiman accused statistics of clinging to elegant data-generating models while the “algorithmic modeling” culture — judged purely on PREDICTIVE accuracy — solved real problems. Prediction and inference are different games with different rules.
Required reading for social scientists using ML. The prediction-versus-causation distinction it sharpened is now the organizing principle of ML-in-economics research.
2001
AnalyticalText Methods
Latent Dirichlet Allocation
Blei, Ng & Jordan
Journal of Machine Learning Research, 2003
Documents as mixtures of latent topics; topics as distributions over words. LDA gave researchers a principled, unsupervised way to discover WHAT texts are about — no dictionaries, no human labels.
The topic-modeling tool behind a wave of accounting and finance papers measuring the content of filings, analyst reports, and conference calls.
2003
2006
EmpiricalNeural Networks
Reducing the Dimensionality of Data with Neural Networks
Hinton & Salakhutdinov
Science, 2006
The comeback paper. Layer-by-layer pretraining made DEEP networks trainable for the first time, beating PCA at compression and shocking a field that had written neural networks off. The phrase “deep learning” entered the mainstream here.
Lit the fuse of the deep-learning decade that followed.
2010s
2012
EmpiricalNeural Networks
ImageNet Classification with Deep Convolutional Neural Networks
Krizhevsky, Sutskever & Hinton
NeurIPS, 2012
AlexNet cut the ImageNet error rate nearly in HALF overnight — GPUs, ReLUs, dropout, and scale. Computer vision converted to deep learning within two years; every other field followed.
The Big Bang of modern AI. The moment 'more data + more compute + deeper networks' became the dominant paradigm.
AnalyticalText Methods
Efficient Estimation of Word Representations in Vector Space
Mikolov, Chen, Corrado & Dean
arXiv / ICLR Workshop, 2013
word2vec: words as dense vectors learned from raw text, where geometry encodes meaning — king − man + woman ≈ queen. Cheap to train on billions of words, and suddenly every NLP system had a semantic foundation.
The moment language became linear algebra. Embeddings are the measurement layer of a decade of social-science text papers — corporate culture, media slant, political risk.
2013
AnalyticalEfficiency & Distillation
Distilling the Knowledge in a Neural Network
Hinton, Vinyals & Dean
NIPS Deep Learning Workshop (arXiv), 2015
The original DISTILLATION paper: train a small, cheap “student” model to match the soft output probabilities of a large “teacher” — and keep most of the accuracy at a fraction of the cost. The teacher's uncertainty, not just its answers, carries the knowledge.
The intellectual basis of every cheap model that punches above its size — from DistilBERT to the small models now running on phones.
2015
AnalyticalArchitectures
Attention Is All You Need
Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser & Polosukhin
NeurIPS, 2017
Threw away recurrence entirely. The TRANSFORMER processes all tokens in parallel, letting attention decide what matters — massively scalable on modern hardware. Eight authors at Google reshaped the entire trajectory of AI in eleven pages.
The architecture of everything: GPT, BERT, Gemini, Claude, LLaMA. Arguably the most consequential ML paper of the century so far.
2017
2019
EmpiricalArchitectures
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, Chang, Lee & Toutanova
NAACL, 2019
Pre-train once on all of Wikipedia by masking words and predicting them; fine-tune cheaply for any task. BERT swept eleven NLP benchmarks and made TRANSFER LEARNING the standard workflow for text.
The model that brought modern NLP to applied research — including the finance- and accounting-specific FinBERTs that followed.
2020s
2020
EmpiricalScaling & Alignment
Language Models are Few-Shot Learners
Brown et al. (OpenAI)
NeurIPS, 2020
GPT-3: 175 billion parameters, trained on much of the internet — and capabilities nobody explicitly programmed. Show it a few examples in the prompt and it performs tasks it was never trained for. IN-CONTEXT LEARNING emerged from scale alone.
The proof that scale creates qualitatively new behavior — and the direct ancestor of ChatGPT.
EmpiricalScaling & Alignment
Scaling Laws for Neural Language Models
Kaplan, McCandlish, Henighan, Brown et al.
Working paper (arXiv), 2020
Model loss falls as a smooth POWER LAW in parameters, data, and compute — predictable across orders of magnitude. Suddenly you could budget intelligence: decide the loss you want, compute the model you need.
Turned AI progress from craft into engineering, and justified the billion-dollar training runs that followed.
AnalyticalEfficiency & Distillation
LoRA: Low-Rank Adaptation of Large Language Models
Hu, Shen, Wallis, Allen-Zhu, Li, Wang, Wang & Chen
ICLR, 2022
The fine-tuning hack that democratized LLM customization: freeze the giant model, train only tiny low-rank matrices injected into each layer — 10,000× fewer trainable parameters, no inference latency, and near-full fine-tuning quality.
Why anyone — including accounting researchers — can now adapt a frontier-class model on a single GPU.
2022
EmpiricalScaling & Alignment
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, Wang, Schuurmans, Bosma, Ichter, Xia, Chi, Le & Zhou
NeurIPS, 2022
Ask the model to “think step by step” and reasoning performance jumps — dramatically, and only above a certain scale. Intermediate steps in plain language turned out to be a computational resource.
Founded the prompting-as-methodology literature and previewed today's reasoning models that deliberate before answering.
EmpiricalScaling & Alignment
Training Language Models to Follow Instructions with Human Feedback
Ouyang et al. (OpenAI)
NeurIPS, 2022
Raw language models complete text; they don't HELP. InstructGPT used reinforcement learning from human feedback (RLHF) to make models follow instructions, refuse harm, and admit uncertainty — with a 1.3B model preferred by humans over a 175B raw one.
The alignment step that made ChatGPT possible. 'Helpful, honest, harmless' became an engineering objective.
EmpiricalEfficiency & Distillation
Training Compute-Optimal Large Language Models
Hoffmann et al. (DeepMind)
NeurIPS, 2022
The “Chinchilla” paper: the field had been building models TOO BIG and training them on too little data. For a fixed compute budget, a 4× smaller model trained on 4× more tokens wins — Chinchilla (70B) beat Gopher (280B) across the board.
Rewrote the economics of training overnight: the cost-optimal frontier is about DATA, not just parameters.
2023
EmpiricalScaling & Alignment
Sparks of Artificial General Intelligence: Early Experiments with GPT-4
Bubeck et al. (Microsoft Research)
Working paper (arXiv), 2023
A 155-page tour of GPT-4 solving problems across math, coding, medicine, law, and psychology — often with no task-specific training. Controversial title, careful documentation. Whatever it is, it is not a stochastic parrot.
The capabilities inventory that made 'general-purpose technology' the default frame for LLMs in economics.
EmpiricalAgents
ReAct: Synergizing Reasoning and Acting in Language Models
Yao, Zhao, Yu, Du, Shafran, Narasimhan & Cao
ICLR, 2023
The recipe behind most LLM AGENTS: interleave chain-of-thought reasoning with actions (search, lookup, tool calls) and feed the observations back into the reasoning loop. Thought → act → observe → think again.
The canonical agent loop — the architecture underneath research assistants, coding agents, and autonomous workflows.
EmpiricalAgents
Generative Agents: Interactive Simulacra of Human Behavior
Park, O'Brien, Cai, Morris, Liang & Bernstein
UIST, 2023
The “Smallville” paper: 25 LLM agents with memory, reflection, and planning living in a simulated town — they spread invitations, formed opinions of each other, and coordinated a Valentine's party nobody scripted. Believable emergent social behavior from language models.
Founded the agent-based-simulation agenda: societies of LLM agents as laboratories for social science.
EmpiricalFinance LLMs
FinBERT: A Large Language Model for Extracting Information from Financial Text
Huang, Wang & Yang
Contemporary Accounting Research, 2023
The accounting profession's FinBERT: pre-trained on 4.9 billion tokens of financial text — 10-Ks, earnings calls, analyst reports. Beats dictionaries and generic BERT at sentiment, ESG classification, and identifying forward-looking statements, especially where financial language diverges from everyday English.
The reference domain-LLM for accounting research, and the standard baseline in the text-analysis literature since.
2025
EmpiricalEfficiency & Distillation
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI
Technical report (arXiv), 2025
The cost shock of the LLM era. Reasoning ability trained with pure reinforcement learning on a mixture-of-experts base whose predecessor (V3) reported a headline training cost near $5.6 million — one to two orders of magnitude below Western frontier runs — with open weights and aggressive distillation into small models.
The week AI-infrastructure stocks repriced. Whatever the true all-in cost, it moved the debate from 'can we afford frontier AI?' to 'how cheap does this get?'
EmpiricalEfficiency & Distillation
DeepSeek-OCR: Contexts Optical Compression
Wei, Sun & Li (DeepSeek-AI)
Technical report (arXiv), 2025
The OCR hack, empirically tested: render long text as an IMAGE and let a vision encoder compress it — the report claims ~97% decoding precision at under 10× compression, meaning ten pages of tokens squeezed into one page's worth of vision tokens. Context windows priced like pictures instead of words.
A technical report, not peer-reviewed — but if the compression numbers hold up, a serious lever on the cost of long-document analysis (10-Ks included).
2010s
2016
EmpiricalBenchmark Design
SQuAD: 100,000+ Questions for Machine Comprehension of Text
Rajpurkar, Zhang, Lopyrev & Liang
EMNLP, 2016
The benchmark that started the leaderboard era of NLP: 100,000 crowd-sourced questions over Wikipedia paragraphs, with exact-match and F1 scoring anyone could reproduce. Within two years, models passed the human baseline — and the field learned its first lesson about benchmark SATURATION.
The template for the modern eval: big dataset, clean metric, public leaderboard — and a lifespan shorter than anyone expected.
2019
EmpiricalBenchmark Design
GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding
Wang, Singh, Michael, Hill, Levy & Bowman
ICLR, 2019
Nine language-understanding tasks bundled into ONE score, so models could no longer overfit a single task. BERT promptly crushed it; SuperGLUE arrived a year later and was crushed too. The eval treadmill was born: build benchmark, saturate benchmark, repeat.
The multi-task evaluation paradigm every LLM leaderboard descends from — and the classic case study in how fast benchmarks die.
2020s
AnalyticalBenchmark Design
Dynabench: Rethinking Benchmarking in NLP
Kiela et al.
NAACL, 2021
The HOW-TO-BUILD-EVALS manifesto: static benchmarks saturate and leak into training data, so collect them DYNAMICALLY — humans adversarially writing examples that current models get wrong, with the model in the loop. The benchmark evolves as the models do.
The intellectual case for living, adversarial evaluation — the antidote to training-set contamination and leaderboard overfitting.
AnalyticalCritiques
AI and the Everything in the Whole Wide World Benchmark
Raji, Bender, Paullada, Denton & Hanna
NeurIPS Datasets & Benchmarks, 2021
The critique every eval builder should read: benchmarks like ImageNet and GLUE get treated as measures of GENERAL capability when they are narrow, arbitrary samples with contested construct validity. A high score means the model is good at the benchmark — everything else is inference.
Construct validity comes to AI: know what your eval actually measures before you claim what the model can do.
2021
EmpiricalKnowledge & Reasoning
Measuring Massive Multitask Language Understanding
Hendrycks, Burns, Basart, Zou, Mazeika, Song & Steinhardt
ICLR, 2021
MMLU: 57 subjects — law, medicine, accounting, philosophy — testing knowledge acquired in pretraining, not task-specific tricks. GPT-3 scored ~44% when random is 25%; frontier models now near-saturate it. For four years it was THE headline number in every model release.
The de facto IQ score of the LLM era. Its saturation is why harder successors (GPQA, MMLU-Pro) exist.
EmpiricalCode & Agents
Evaluating Large Language Models Trained on Code
Chen et al. (OpenAI)
Working paper (arXiv), 2021
The Codex paper, and with it HumanEval: hand-WRITTEN programming problems (so they can't be memorized from GitHub) scored by whether generated code passes unit tests — the pass@k metric. Functional correctness replaced textual similarity as the standard for code.
Defined how the world measures AI coding — execution-based evaluation, not string matching — and launched the tool now called GitHub Copilot.
EmpiricalKnowledge & Reasoning
Training Verifiers to Solve Math Word Problems
Cobbe et al. (OpenAI)
Working paper (arXiv), 2021
GSM8K: 8,500 grade-school math word problems that stumped GPT-3 — plus the idea of training VERIFIERS to judge candidate solutions and rerank them, foreshadowing today's reasoning models scored by outcome checks.
The standard math-reasoning eval for half a decade, and an early blueprint for verification-based training.
EmpiricalFinance Evals
FinQA: A Dataset of Numerical Reasoning over Financial Data
Chen, Chen, Smiley, Shah et al.
EMNLP, 2021
The finance-domain eval: 8,281 expert-written questions requiring MULTI-STEP numerical reasoning over real earnings-report tables and text — compute the ratio, compare the years, chain the operations. Models must show an executable reasoning program, not just an answer.
The benchmark for whether AI can actually do financial-statement work — and the ancestor of every accounting-domain eval since.
2022
EmpiricalTruthfulness & Safety
TruthfulQA: Measuring How Models Mimic Human Falsehoods
Lin, Hilton & Evans
ACL, 2022
The hallucination eval with a twist: 817 questions where the TRAINING DATA itself is wrong — misconceptions, conspiracy theories, old wives' tales. Bigger models were initially LESS truthful, because scale means better imitation of human text, errors included.
Formalized truthfulness as measurable and separable from capability — the conceptual basis of every hallucination benchmark.
AnalyticalBenchmark Design
Holistic Evaluation of Language Models
Liang et al. (Stanford CRFM)
Transactions on Machine Learning Research, 2023
HELM: the framework paper for building eval SUITES. Taxonomize scenarios (tasks × domains × users) and metrics (accuracy, calibration, robustness, fairness, bias, toxicity, efficiency), then measure every model on every cell — exposing that models topping accuracy leaderboards often fail on calibration or robustness.
The methodology reference for anyone designing an evaluation: multi-metric, transparent, and explicit about what is NOT covered.
2023
EmpiricalBenchmark Design
Beyond the Imitation Game: Quantifying and Extrapolating the Capabilities of Language Models
Srivastava et al. (442 authors)
Transactions on Machine Learning Research, 2023
BIG-bench: 204 tasks crowd-sourced from 132 institutions — from chess to social bias to ancient languages — built to measure capabilities BEYOND saturated benchmarks and to study how performance scales, including the 'emergent' jumps that appear only above certain model sizes.
The community-scale approach to eval construction, and a primary data source for the emergent-abilities debate.
EmpiricalKnowledge & Reasoning
GPQA: A Graduate-Level Google-Proof Q&A Benchmark
Rein et al.
Working paper (arXiv), 2023
The eval built for the post-MMLU world: PhD-written science questions so hard that skilled non-experts WITH GOOGLE score only ~34% (experts ~65%). Designed explicitly for scalable oversight — how do you evaluate a model on questions you can't answer yourself?
The current frontier knowledge benchmark — and a preview of the deepest eval problem: judging systems smarter than their judges.
EmpiricalJudging & Arenas
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Zheng, Chiang, Sheng et al.
NeurIPS Datasets & Benchmarks, 2023
The two evals that rule the chatbot era: MT-Bench (multi-turn questions scored by GPT-4) and Chatbot Arena (crowdsourced blind A/B battles aggregated into Elo ratings). Crucially, it VALIDATED the judge: strong LLM judges agree with humans ~80% — about as often as humans agree with each other — while documenting position, verbosity, and self-preference biases.
Made LLM-as-a-judge a legitimate, audited methodology — the scalable evaluation trick behind most modern model comparisons.
2024
EmpiricalCode & Agents
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
The AGENTIC eval: 2,294 real GitHub issues from real repositories — fix the bug so the actual test suite passes, navigating a full codebase, not a toy function. Frontier models initially solved under 2%; agent scaffolds now push far higher, making it the yardstick of the coding-agent race.
The benchmark that moved evaluation from answering questions to DOING WORK — the standard for measuring autonomous AI agents.
2010s
AnalyticalInference with ML
High-Dimensional Methods and Inference on Structural and Treatment Effects
Belloni, Chernozhukov & Hansen
Journal of Economic Perspectives, 2014
When you have thousands of potential controls, LASSO can pick them — but naïve selection breaks inference. Post-double-selection fixes it: select variables for BOTH the outcome and the treatment equations, and standard errors become honest again.
The paper that taught economists you can use ML for variable selection without sacrificing valid confidence intervals.
2014
AnalyticalPrediction Policy
Prediction Policy Problems
Kleinberg, Ludwig, Mullainathan & Obermeyer
American Economic Review P&P, 2015
The agenda-setting distinction: some policy problems are CAUSAL (does the treatment work?) but many are PURE PREDICTION (who will it rain on? which hip replacement patient will die within a year?). For prediction problems, ML — not econometrics — is the right tool, and the payoffs are large.
Four pages that created a research program: find the decisions where better prediction alone creates value.
2015
2016
EmpiricalBias & Fairness
Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
Bolukbasi, Chang, Zou, Saligrama & Kalai
NeurIPS, 2016
Word embeddings trained on news text learn the analogies of our prejudices — literally completing 'man : programmer :: woman : homemaker.' The paper both DOCUMENTED the bias geometrically and proposed the first debiasing algorithms.
The paper that made 'AI inherits our biases' a measurable, fixable, and unavoidable research question.
AnalyticalPrediction vs Causation
Machine Learning: An Applied Econometric Approach
Mullainathan & Spiess
Journal of Economic Perspectives, 2017
The clearest statement of what ML is FOR in economics: ML solves y-hat problems (prediction), econometrics solves beta-hat problems (parameters). Off-the-shelf ML predictions are excellent; naïvely reading coefficients off ML models is malpractice — regularization makes them uninterpretable.
The y-hat versus beta-hat distinction is now the first lecture of every ML-for-economists course.
AnalyticalPrediction Policy
Beyond Prediction: Using Big Data for Policy Problems
Athey
Science, 2017
The roadmap for combining the two cultures: prediction models allocate attention and resources; causal models evaluate interventions; the interesting frontier is where they meet — heterogeneous treatment effects, policy targeting, and bandit-style learning.
The synthesis piece that framed the ML-meets-causal-inference agenda for the decade.
2017
EmpiricalBias & Fairness
Semantics Derived Automatically from Language Corpora Contain Human-Like Biases
Caliskan, Bryson & Narayanan
Science, 2017
Replicated the Implicit Association Test ON A MACHINE: embeddings reproduce the full catalogue of documented human implicit biases — flowers vs. insects, race and names, gender and careers — with effect sizes matching the psychology literature.
The cleanest demonstration that bias lives in the training corpus itself: any model trained on human text starts with human prejudice.
AnalyticalInference with ML
Estimation and Inference of Heterogeneous Treatment Effects using Random Forests
Wager & Athey
Journal of the American Statistical Association, 2018
CAUSAL FORESTS: random forests rebuilt to estimate treatment-effect heterogeneity, with asymptotic normality and valid confidence intervals. Adaptive nearest-neighbor matching, but with statistical guarantees.
The workhorse for 'for whom does it work?' questions — from job training to disclosure regulation.
AnalyticalInference with ML
Double/Debiased Machine Learning for Treatment and Structural Parameters
Use ANY ML method to soak up high-dimensional nuisance functions, then estimate the low-dimensional parameter you actually care about — with Neyman orthogonality and cross-fitting protecting the inference from the ML's biases.
The general license to put deep learning inside a causal estimate. DML is now standard in applied micro, finance, and accounting.
2018
EmpiricalMan vs Machine
Human Decisions and Machine Predictions
Kleinberg, Lakkaraju, Leskovec, Ludwig & Mullainathan
Quarterly Journal of Economics, 2018
The landmark man-versus-machine comparison: judges deciding bail versus an ML model predicting flight risk. The machine's ranking dominates — crime could fall ~25% with no increase in jailing, or jail populations could fall ~40% with no increase in crime. Judges systematically mispredict, and the paper shows exactly where.
THE paper comparing algorithmic prediction with human learning and judgment — and the methodological playbook for doing that comparison honestly (selective labels and all).
AnalyticalPrediction vs Causation
Machine Learning Methods That Economists Should Know About
Athey & Imbens
Annual Review of Economics, 2019
The consolidated toolkit: regularized regression, trees and forests, matrix completion, and the causal adaptations of each — written by the two economists who did the most to translate ML into econometrics.
The standard reference syllabus for the ML-econometrics interface.
2019
EmpiricalBias & Fairness
Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations
Obermeyer, Powers, Vogeli & Mullainathan
Science, 2019
A deployed algorithm guiding care for millions used health COSTS as the label for health NEEDS — and because less money is spent on Black patients, the algorithm systematically understated their illness. Fixing the label nearly tripled the share of Black patients flagged for extra care.
The canonical demonstration that algorithmic bias is usually a LABEL-CHOICE problem — an economics problem — not a math problem.
2020s
2022
EmpiricalBias & Fairness
Predictably Unequal? The Effects of Machine Learning on Credit Markets
Fuster, Goldsmith-Pinkham, Ramadorai & Walther
Journal of Finance, 2022
Better prediction technology does not lift all boats: moving from logit to ML in mortgage screening produces WINNERS AND LOSERS, and the losers concentrate among minority borrowers — not from biased data, but from the geometry of flexible models on unequal populations.
The distributional economics of prediction: accuracy and equity can move in opposite directions even with clean data.
EmpiricalMan vs Machine
Diagnosing Physician Error: A Machine Learning Approach to Low-Value Health Care
Mullainathan & Obermeyer
Quarterly Journal of Economics, 2022
Compared physicians' heart-attack testing decisions against an ML benchmark of true risk: doctors OVER-test low-risk patients and UNDER-test high-risk ones — errors traced to simple heuristics and salient symptoms. The algorithm reveals the structure of expert misjudgment.
The mature version of man-versus-machine research: use the model not to replace the expert but to DIAGNOSE how the expert learns and errs.
2023
EmpiricalLook-Ahead & Leakage
Assessing Look-Ahead Bias in Stock Return Predictions Generated by GPT Sentiment Analysis
Glasserman & Lin
Working paper (Columbia), 2023
The uncomfortable question for every 'ChatGPT predicts returns' paper: the model may simply REMEMBER what happened. Using anonymized headlines and comparing periods inside versus outside the training window, they find look-ahead bias and distraction effects contaminating naive LLM return predictions.
The methodological cold shower for LLM-based asset pricing: if the future is in the training data, your alpha is memory.
2024
EmpiricalLook-Ahead & Leakage
Lookahead Bias in Pretrained Language Models
Sarkar & Vafa
Working paper (SSRN), 2024
Shows that pretrained LLMs leak post-dated knowledge even when prompted to role-play an earlier date — models 'know' about events after the claimed forecast date — and proposes evaluation protocols and time-partitioned models for valid backtesting.
Why serious LLM finance research now trains or evaluates on time-sliced corpora. Leakage is the new p-hacking.
AnalyticalLook-Ahead & Leakage
Large Language Models: An Applied Econometric Framework
Ludwig, Mullainathan & Rambachan
NBER Working Paper No. 33344, 2025
The econometrics of using LLMs in research: when can LLM-generated data stand in for human responses, and when does training-data leakage invalidate the design? Provides conditions — no contamination between training corpus and estimation sample — under which LLM-based measurement supports valid inference.
The formal rulebook the field was missing: treat the LLM as a measurement instrument with an unknown training set, and design around it.
2025
2000s
2004
EmpiricalSentiment & Tone
Is All That Talk Just Noise? The Information Content of Internet Stock Message Boards
Antweiler & Frank
Journal of Finance, 2004
The first serious NLP on investor social media: 1.5 million Yahoo! Finance and Raging Bull messages, classified by machine. Message-board talk predicted VOLATILITY (not returns), and disagreement predicted trading volume. “Noise” turned out to contain measurable signal.
Twenty years before FinTwit and StockTwits research, this paper set the template: classify the chatter, then test what it moves.
2007
EmpiricalSentiment & Tone
Giving Content to Investor Sentiment: The Role of Media in the Stock Market
Tetlock
Journal of Finance, 2007
Ran the Wall Street Journal's “Abreast of the Market” column through a word-count dictionary. High pessimism predicted downward price pressure followed by REVERSAL — media tone was sentiment, not just information. The paper that made textual analysis respectable in top finance journals.
The founding paper of media-and-markets research. Its design — text measure, then horse-race against fundamentals — is still the standard.
2008
EmpiricalReadability
Annual Report Readability, Current Earnings, and Earnings Persistence
Li
Journal of Accounting and Economics, 2008
Counted syllables and sentence lengths in 55,000 10-Ks. Firms with LONGER, harder-to-read reports have lower earnings and less persistent profits — consistent with managers obfuscating bad news. Readability became a strategic disclosure choice, not a style preference.
Launched the readability literature in accounting and put linguistic obfuscation on the research map.
EmpiricalSentiment & Tone
More Than Words: Quantifying Language to Measure Firms' Fundamentals
Tetlock, Saar-Tsechansky & Macskassy
Journal of Finance, 2008
Negative words in firm-specific news predict LOW earnings — and the market underreacts to that verbal information. Language contains hard fundamental content that numbers haven't captured yet.
Established that text predicts fundamentals, not just sentiment — the intellectual license for two decades of NLP in capital markets research.
2010s
2010
EmpiricalMeasurement
What Drives Media Slant? Evidence from U.S. Daily Newspapers
Gentzkow & Shapiro
Econometrica, 2010
Measured each newspaper's political slant by matching its language to phrases distinctive of congressional Republicans versus Democrats. The punchline is economic: slant tracks what the paper's READERS want, not what its owner believes — media bias is demand-driven.
The founding text-as-data paper in economics: language as measurement, then economics done with the measure.
EmpiricalMeasurement
The Information Content of Forward-Looking Statements in Corporate Filings — A Naïve Bayesian Machine Learning Approach
Li
Journal of Accounting Research, 2010
One of the FIRST supervised machine-learning applications in accounting: a naïve Bayes classifier categorizing 30 million forward-looking sentences from MD&As by tone and content. Machine-classified tone predicted future earnings; dictionaries could not match it.
The proof of concept that supervised ML beats word lists for accounting text — a decade before that became consensus.
2011
EmpiricalSentiment & Tone
When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks
Loughran & McDonald
Journal of Finance, 2011
Three-quarters of “negative” words from the standard psychology dictionary are NOT negative in finance — liability, cost, tax, board. Built finance-specific word lists from scratch and showed generic dictionaries systematically mismeasure financial text.
The most-used dictionary in accounting and finance NLP. Also the field's canonical lesson: domain context is everything in text measurement.
2012
EmpiricalDeception
Detecting Deceptive Discussions in Conference Calls
Larcker & Zakolyukina
Journal of Accounting Research, 2012
Trained classifiers on the calls of firms that later restated earnings. Deceptive executives use more references to general knowledge, fewer self-references, and fewer nonextreme positive words — linguistic fingerprints of deception that beat chance by a wide margin.
Founded computational deception detection in accounting. The ancestor of every 'can AI catch lying managers?' paper.
EmpiricalDeception
The Power of Voice: Managerial Affective States and Future Firm Performance
Mayew & Venkatachalam
Journal of Finance, 2012
Ran earnings-call AUDIO through vocal-emotion software. Managers' vocal affect — beyond the words themselves — predicts future performance, analyst forecast errors, and subsequent negative news. The market hears the words but underprices the voice.
Extended textual analysis into paralinguistics: HOW managers speak carries information the transcript misses.
2014
EmpiricalMeasurement
Evidence on the Information Content of Text in Analyst Reports
Huang, Zang & Zheng
The Accounting Review, 2014
Naïve Bayes over 363,952 analyst reports: the TEXT carries information beyond the earnings forecasts and recommendations — and the market reaction to text is stronger when it disagrees with the quantitative summary measures.
Demonstrated that analysts' words are a distinct information channel, not decoration around the numbers.
AnalyticalSurvey
Textual Analysis in Accounting and Finance: A Survey
Loughran & McDonald
Journal of Accounting Research, 2016
The field guide: readability, tone, similarity, and the methodological landmines — document length effects, dictionary misuse, and the perils of overfitting text. Set the standards the literature still cites.
The survey every PhD student reads before touching a 10-K corpus.
2016
EmpiricalMeasurement
Measuring Economic Policy Uncertainty
Baker, Bloom & Davis
Quarterly Journal of Economics, 2016
Counted newspaper articles pairing 'economy', 'policy', and 'uncertainty' into the EPU index — validated against human readings, replicated in a dozen countries, and shown to predict investment and employment dynamics.
The most successful text-based measure in economics: cited tens of thousands of times and updated monthly at policyuncertainty.com.
EmpiricalMeasurement
Text-Based Network Industries and Endogenous Product Differentiation
Hoberg & Phillips
Journal of Political Economy, 2016
Rebuilt industry classification from the product descriptions firms write in their own 10-Ks. Text-based networks capture competition SIC and NAICS codes miss — and they move as firms reposition their products.
Replaced static industry codes with living, text-defined product markets. A landmark of measurement-by-text.
2017
EmpiricalMeasurement
News Implied Volatility and Disaster Concerns
Manela & Moreira
Journal of Financial Economics, 2017
Built a VIX going back to 1890 — from newspaper front pages. Support vector regression maps Wall Street Journal text to implied volatility, extending a key state variable a century before options markets existed.
Text as a TIME MACHINE: machine learning recovering unobservable economic quantities from historical archives.
2018
EmpiricalMeasurement
Analyst Information Discovery and Interpretation Roles: A Topic Modeling Approach
Huang, Lehavy, Zang & Zheng
Management Science, 2018
LDA over analyst reports and conference calls separates what analysts DISCOVER (topics beyond the call) from what they INTERPRET (topics echoing management). Both roles carry price-relevant information — but in different proportions across the information cycle.
Showcased topic models answering a real economic question: what exactly do information intermediaries add?
EmpiricalMeasurement
Transparency and Deliberation Within the FOMC: A Computational Linguistics Approach
Hansen, McMahon & Prat
Quarterly Journal of Economics, 2018
Topic-modeled twenty years of FOMC transcripts around a natural experiment — the 1993 decision to publish them. Transparency changed HOW policymakers deliberate: more discipline and preparation, but also more conformity and scripted discussion.
Machine reading of institutional deliberation: text methods answering a core political-economy question about transparency itself.
AnalyticalSurvey
Text as Data
Gentzkow, Kelly & Taddy
Journal of Economic Literature, 2019
The methodological synthesis: text as high-dimensional data, the representation choices (counts, embeddings, topics), and the inference pitfalls when features number in the millions. Written by economists, for economists.
The canonical bridge between NLP and econometrics — the survey that made text methods part of the economist's standard toolkit.
2019
EmpiricalCulture & Risk
Firm-Level Political Risk: Measurement and Effects
Hassan, Hollander, van Lent & Tahoun
Quarterly Journal of Economics, 2019
Measured each firm's political risk from the political share of its earnings-call language. The striking finding: most variation in political risk is FIRM-level, not aggregate — and firms facing it hedge, lobby, and retrench investment.
The template for firm-level 'exposure' measures mined from calls — political, climate, pandemic, and beyond.
2020s
2021
EmpiricalCulture & Risk
Measuring Corporate Culture Using Machine Learning
Li, Mai, Shen & Yan
Review of Financial Studies, 2021
Word embeddings trained on 200,000+ earnings calls score firms on the 'Big Five' cultural values — innovation, integrity, quality, respect, teamwork. Culture scores predict business outcomes and shift measurably after shocks like fraud revelations.
Made 'soft' corporate culture a hard, replicable variable — embeddings as measurement instruments.
2022
EmpiricalSentiment & Tone
Disclosure Sentiment: Machine Learning vs. Dictionary Methods
Frankel, Jennings & Lee
Management Science, 2022
The head-to-head: machine learning extracts sentiment from disclosures that explains market reactions BETTER than Loughran-McDonald dictionaries — but the gains are largest where dictionaries are weakest, and interpretability trades off against power.
The measured verdict on ML-versus-dictionaries, and a model of how to benchmark text methods honestly.
2023
EmpiricalMachine Readers
How to Talk When a Machine Is Listening: Corporate Disclosure in the Age of AI
Cao, Jiang, Yang & Zhang
Review of Financial Studies, 2023
Firms KNOW the machines are reading. As machine ownership of their shares rises, companies rewrite filings to be machine-friendly — avoiding words that algorithms score as negative and adjusting tone the machines will parse. Disclosure is now written for two audiences.
The feedback-loop paper: AI doesn't just read disclosure, it CHANGES it. Directly relevant to strategic-communication research.
2024
EmpiricalMeasurement
Business News and Business Cycles
Bybee, Kelly, Manela & Xiu
Journal of Finance, 2024
Topic-modeled the full text of the Wall Street Journal, 1984–2017, into 180 interpretable news themes whose attention shares track and FORECAST macro activity — a narrative decomposition of the business cycle.
Text as macro data: the news is a real-time survey of the economy, and machine-measured attention is a state variable.
2025
EmpiricalCulture & Risk
Beyond Words: The Challenge of Interpreting Financial Text Across Cultures
Pinto & Siano
Journal of International Business Studies, 2025
US-trained NLP tools stumble abroad: the same disclosure language carries different meaning across cultural and linguistic contexts, and standard US-based approaches systematically mismeasure tone in international filings.
A caution for the globalization of text-based research: language models inherit the culture of their training data.
1960s
1968
EmpiricalCredit & Lending
Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy
Altman
Journal of Finance, 1968
The Z-score: five accounting ratios, one discriminant function, and bankruptcy prediction accuracy that stunned the field. The first famous demonstration that statistical models beat expert judgment on a core financial task.
Still quoted on trading desks six decades later. The ancestor of every credit-risk model — statistical or neural.
1980s
1980
EmpiricalCredit & Lending
Financial Ratios and the Probabilistic Prediction of Bankruptcy
Ohlson
Journal of Accounting Research, 1980
Replaced discriminant analysis with conditional LOGIT — probabilities instead of classifications, and a fix for the methodological problems in matched-sample designs. The O-score.
Methodological rigor arriving in prediction research; the workhorse specification for decades of distress studies.
1989
EmpiricalEarnings Prediction
Financial Statement Analysis and the Prediction of Stock Returns
Ou & Penman
Journal of Accounting and Economics, 1989
Fed 68 accounting ratios into a statistical model to produce Pr — the probability that earnings rise next year — and showed that trading on it earned abnormal returns. Fundamental analysis, mechanized.
The intellectual grandfather of every machine-learning-on-fundamentals paper: the information is in the statements, if you have the machinery to extract it.
2010s
2011
EmpiricalFraud & Misstatement
Predicting Material Accounting Misstatements
Dechow, Ge, Larson & Sloan
Contemporary Accounting Research, 2011
Built the F-SCORE from a comprehensive database of SEC enforcement actions: a scaled probability of misstatement from accruals quality, performance, and market signals. A red-flag index practitioners actually adopted.
The benchmark every ML fraud model has to beat — and the paper that defined the misstatement-prediction task.
EmpiricalFraud & Misstatement
Financial Statement Fraud Detection: An Analysis of Statistical and Machine Learning Algorithms
Perols
Auditing: A Journal of Practice & Theory, 2011
The first systematic horse-race of ML classifiers — SVMs, neural networks, logistic regression, bagging — on fraud detection, with careful attention to the extreme class imbalance that makes fraud prediction statistically brutal.
Early, honest evidence on what ML can and cannot add when positives are rare — a lesson the field keeps relearning.
2020s
2020
EmpiricalFraud & Misstatement
Detecting Accounting Fraud in Publicly Traded U.S. Firms Using a Machine Learning Approach
Bao, Ke, Li, Yu & Zhang
Journal of Accounting Research, 2020
Ensemble learning on RAW financial statement numbers — no researcher-designed ratios — beats the F-score and logit benchmarks out of sample. The machine finds the ratios itself.
The JAR paper that legitimized pure-ML prediction in top-tier accounting research, and open-sourced its fraud data for the field.
EmpiricalCredit & Lending
Machine + Man: A Field Experiment on the Role of Discretion in Augmenting AI-Based Lending Models
Costello, Down & Mehta
Journal of Accounting and Economics, 2020
A REAL field experiment at a commercial lender: give loan officers discretion to override the machine's credit scores. Human overrides underperform the model on average — but discretion helps exactly where the machine's data is thinnest.
One of the cleanest tests of human-plus-machine decision-making in accounting: the answer is a division of labor, not a winner.
2021
EmpiricalFraud & Misstatement
Using Machine Learning to Detect Misstatements
Bertomeu, Cheynel, Floyd & Pan
Review of Accounting Studies, 2021
Gradient-boosted trees over accounting, audit, and market variables. Beyond accuracy, the paper reads the model's feature importances ECONOMICALLY: which signals matter, when, and why — turning a black box into evidence about the anatomy of misstatement.
A model for interpretable ML in accounting: prediction as a microscope, not just a scorecard.
EmpiricalGovernance
Selecting Directors Using Machine Learning
Erel, Stern, Tan & Weisbach
Review of Financial Studies, 2021
Trained on director histories, the algorithm predicts which board appointees will underperform — and the directors the model would NOT have chosen are more likely to be friends of the CEO, male, and well-connected. The machine exposes the favoritism in the boardroom.
Algorithmic decision aids as governance diagnostics: the gap between model choice and human choice IS the agency problem, quantified.
2022
EmpiricalAnalysts vs AI
Human Versus Machine: A Comparison of Robo-Analyst and Traditional Research Analyst Investment Recommendations
Coleman, Merkley & Pacelli
The Accounting Review, 2022
Robo-analysts — automated research platforms — issue more balanced recommendations than human analysts, revise on filings rather than headlines, and their buys earn abnormal returns. Fewer conflicts, different information diet, real performance.
First systematic evidence on automated equity research: the machines are not just cheaper, they are differently biased.
EmpiricalEarnings Prediction
Predicting Future Earnings Changes Using Machine Learning and Detailed Financial Data
Chen, Cho, Dou & Lev
Journal of Accounting Research, 2022
Machine learning over the FULL detail of financial statements — hundreds of granular line items, not summary ratios — predicts earnings changes better than analysts and generates profitable trading signals. The detail the literature aggregated away was the signal.
Vindicated the accounting system itself: the value is in the granularity, and ML is the technology that finally exploits it.
2023
EmpiricalAnalysts vs AI
Man versus Machine Learning: The Term Structure of Earnings Expectations and Conditional Biases
van Binsbergen, Han & Lopez-Lira
Review of Financial Studies, 2023
Built a machine-learning benchmark for expected earnings and measured analysts AGAINST it: analyst forecasts are predictably biased — too optimistic at long horizons — and the machine-implied expectation errors predict returns.
Turned ML into an econometric instrument: a rational benchmark against which human belief distortions become measurable.
2024
EmpiricalAnalysts vs AI
From Man vs. Machine to Man + Machine: The Art and AI of Stock Analyses
Cao, Jiang, Wang & Yang
Journal of Financial Economics, 2024
AI analysts beat most human analysts — but humans WIN where soft information and institutional knowledge matter: illiquid small firms, asset-light business models, distress. And man + machine beats either alone.
The definitive comparative-advantage map for AI in investment research: substitution where data is rich, complementarity where judgment is scarce.
2010s
AnalyticalFactor Zoo
Presidential Address: Discount Rates
Cochrane
Journal of Finance, 2011
The address that named the problem: a “ZOO of new factors” — hundreds of characteristics claimed to predict returns — with no discipline on which provide INDEPENDENT information. Cochrane reframed asset pricing around discount-rate variation and issued the challenge the next decade of research answered.
The origin of the term 'factor zoo' and the research agenda of taming it — the question every factor-selection paper cites in its first paragraph.
2011
2016
EmpiricalFactor Zoo
… and the Cross-Section of Expected Returns
Harvey, Liu & Zhu
Review of Financial Studies, 2016
Catalogued 316 published factors and dropped the multiple-testing bomb: after adjusting for the sheer number of tests the profession has run, the usual t-statistic of 2.0 is worthless — new factors should clear a hurdle of roughly t > 3.0. Most published factors are likely FALSE discoveries.
Changed empirical finance's evidentiary standards overnight. The reason referees now ask about data snooping by default.
EmpiricalFactor Zoo
Does Academic Research Destroy Stock Return Predictability?
McLean & Pontiff
Journal of Finance, 2016
Tracked 97 published predictors OUT of sample: returns decay ~26% after the sample ends and ~58% after PUBLICATION — with the decay concentrated in liquid, arbitrageable stocks. Investors read the journals and trade the anomalies away.
The elegant test separating statistical bias from real-but-arbitraged predictability: publication itself is the treatment.
2019
EmpiricalFactor Models
Characteristics Are Covariances: A Unified Model of Risk and Return
Kelly, Pruitt & Su
Journal of Financial Economics, 2019
Instrumented PCA lets factor loadings depend on firm characteristics — collapsing the zoo of anomaly characteristics into a handful of latent risk factors estimated from the data. Characteristics predict returns BECAUSE they proxy for time-varying betas.
Modern statistical machinery applied to the oldest question in asset pricing: risk or mispricing?
2020s
2020
EmpiricalFactor Zoo
Taming the Factor Zoo: A Test of New Factors
Feng, Giglio & Xiu
Journal of Finance, 2020
THE selection machine: double-selection LASSO evaluates any NEW factor against the entire existing zoo, asking whether it adds pricing information beyond what hundreds of incumbents already span — with valid inference despite the high-dimensional controls. Verdict: most recent factors are redundant; only a handful earn their keep.
Exactly the protocol the field needed — a disciplined, ML-based answer to 'does your factor survive the zoo?' Now the standard hurdle for proposing one.
EmpiricalFactor Zoo
Replicating Anomalies
Hou, Xue & Zhang
Review of Financial Studies, 2020
The great replication purge: re-tested 452 published anomalies with standard controls — NYSE breakpoints, value-weighted returns, microcaps handled properly — and 65% FAIL at conventional significance. With the higher t-hurdle, ~82% fail. Much of the zoo was microcap noise and equal-weighting.
The empirical housecleaning of the cross-section. Read alongside Jensen-Kelly-Pedersen for the full debate.
EmpiricalReturn Prediction
Empirical Asset Pricing via Machine Learning
Gu, Kelly & Xiu
Review of Financial Studies, 2020
The benchmark study: every major ML method — trees, forests, boosting, neural nets — horse-raced on predicting US stock returns with 94 characteristics. Neural networks and trees roughly DOUBLE the out-of-sample R² of regressions, mostly by capturing interactions no linear model sees.
The paper that made machine learning a standard tool in empirical asset pricing. The methodological template for hundreds of follow-ups.
EmpiricalReturn Prediction
Dissecting Characteristics Nonparametrically
Freyberger, Neuhierl & Weber
Review of Financial Studies, 2020
Adaptive group LASSO asks two questions the anomalies literature dodged: which characteristics matter JOINTLY (only ~a dozen of 62 survive), and is their effect nonlinear (emphatically yes).
Variable selection with discipline: most of the anomaly zoo is redundant once you let a few characteristics act nonlinearly.
EmpiricalFactor Models
Shrinking the Cross-Section
Kozak, Nagel & Santosh
Journal of Financial Economics, 2020
Bayesian shrinkage over thousands of characteristic portfolios constructs a robust stochastic discount factor — regularization as economics, since absence of near-arbitrage implies the SDF loads on a few dominant principal components.
Machine-learning regularization, derived from asset-pricing theory rather than borrowed from computer science.
2021
EmpiricalBonds
Bond Risk Premiums with Machine Learning
Bianchi, Büchner & Tamoni
Review of Financial Studies, 2021
Extended the ML toolkit to Treasury markets: neural networks uncover bond-return predictability from macro and yield-curve information that spanning regressions miss, with economic gains for bond investors.
Evidence that the ML advantage generalizes beyond equities — nonlinearity is a feature of asset markets, not a quirk of stocks.
2023
EmpiricalFactor Zoo
Is There a Replication Crisis in Finance?
Jensen, Kelly & Pedersen
Journal of Finance, 2023
The rebuttal: a Bayesian hierarchical model that shares information ACROSS factors — plus a unified global dataset of 153 factors in 93 countries — finds most factors DO replicate, survive out-of-sample, and strengthen when modeled jointly. The zoo is more disciplined than the purge suggested.
The optimistic pole of the replication debate — and a masterclass in using hierarchical ML-style shrinkage to settle a meta-scientific fight.
AnalyticalComplexity
The Virtue of Complexity in Return Prediction
Kelly, Malamud & Zhou
Journal of Finance, 2024
Against ALL conventional wisdom: theory and evidence that bigger models keep helping even when parameters vastly outnumber observations. In the 'double descent' regime, ridge-regularized complexity improves out-of-sample Sharpe ratios — parsimony was the bias.
The theoretical license for large models in finance. The bias-variance folklore that governed empirical finance for decades, overturned.
2024
EmpiricalReturn Prediction
Deep Learning in Asset Pricing
Chen, Pelger & Zhu
Management Science, 2024
Imposed the no-arbitrage moment condition directly on a deep network — GANs estimating the SDF, LSTMs digesting macro time series — delivering an economically disciplined deep model that prices the cross-section out of sample.
The synthesis: not ML instead of theory, but ML constrained BY theory. Where the asset-pricing-meets-AI literature is heading.
1980s
AnalyticalInformation Theory
On the Impossibility of Informationally Efficient Markets
Grossman & Stiglitz
American Economic Review, 1980
The paradox that governs all information technology in markets: if prices fully reveal information, nobody pays to gather it — so equilibrium requires prices to be INEFFICIENT enough to reward the informed. Information acquisition is an economic activity with costs and returns.
The framework for every question about AI in markets: cheaper information processing shifts the equilibrium amount of informed trading — it cannot make markets fully efficient.
1980
2000s
AnalyticalInformation Theory
Implications of Rational Inattention
Sims
Journal of Monetary Economics, 2003
Modeled humans as Shannon channels: finite information-processing CAPACITY, allocated optimally. Rational inattention explains sluggish, discrete responses to news without assuming irrationality — attention is the scarce resource.
The economics of limited attention that AI directly disrupts: machines relax the very constraint this theory is built on.
2003
2010s
AnalyticalAutomation & Labor
Why Are There Still So Many Jobs? The History and Future of Workplace Automation
Autor
Journal of Economic Perspectives, 2015
Two centuries of automation anxiety, and employment keeps growing — because automation COMPLEMENTS the tasks it does not replace, raises incomes that create new demand, and because 'tacit' human tasks resist codification longer than engineers expect (Polanyi's paradox).
The essential base rate for every AI-and-jobs conversation: displacement is real, but so is the two-hundred-year record of task creation.
2015
2017
EmpiricalAutomation & Labor
The Future of Employment: How Susceptible Are Jobs to Computerisation?
Frey & Osborne
Technological Forecasting and Social Change, 2017
The famous '47% of US employment at risk' study: occupation-level automatability scored by ML classifiers trained on expert labels. Widely criticized — task-level studies cut the number dramatically — but it set the terms of the global debate.
The most-cited, most-argued-with number in the automation literature. Read it WITH its critics; that dialogue is the field.
AnalyticalData Economics
Big Data in Finance and the Growth of Large Firms
Begenau, Farboodi & Veldkamp
Journal of Monetary Economics, 2018
Big data is not neutral: large firms generate more data, data attracts cheaper capital, cheaper capital makes them larger. A theory of how the data economy FEEDS scale — and quietly disadvantages young, small, data-poor firms.
The data-feedback-loop model underlying debates about AI, market concentration, and who benefits from the information revolution.
AnalyticalAutomation & Labor
The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment
Acemoglu & Restrepo
American Economic Review, 2018
The task framework: automation DISPLACES labor from existing tasks while technology also CREATES new tasks where labor has comparative advantage. The race between displacement and reinstatement determines wages and labor share — and nothing guarantees the race is won.
The canonical model for thinking about AI and jobs — the reason 'what tasks?' replaced 'how many jobs?' as the right question.
2018
AnalyticalGrowth
Artificial Intelligence and Economic Growth
Aghion, Jones & Jones
In: The Economics of Artificial Intelligence (NBER / U. Chicago Press), 2019
If AI automates the production of IDEAS themselves, growth theory changes character — with Baumol's cost disease as the crucial brake: growth ends up constrained not by what we automate, but by what we CANNOT.
The framework for the biggest question in economics right now: does AI accelerate growth modestly, dramatically, or explosively?
AnalyticalAutomation & Labor
Exploring the Impact of Artificial Intelligence: Prediction versus Judgment
Agrawal, Gans & Goldfarb
Information Economics and Policy, 2019
The clean decomposition: modern AI is a drop in the cost of PREDICTION. Prediction and judgment are complements in decisions — so cheap machine prediction raises the value of human judgment about payoffs, ethics, and objectives.
The single most useful sentence for thinking about AI adoption: machines predict, humans decide what matters.
2019
2020s
AnalyticalData Economics
Long-Run Growth of Financial Data Technology
Farboodi & Veldkamp
American Economic Review, 2020
Why did markets shift from fundamentals research to order-flow analysis? A growth model of data technology: as processing improves, investors rationally reallocate from long-run fundamental information to short-run demand information. Price informativeness about fundamentals need not improve.
The unsettling equilibrium answer to 'does better data technology make prices more informative?' — not necessarily, and maybe the opposite.
2020
EmpiricalAutomation & Labor
Robots and Jobs: Evidence from US Labor Markets
Acemoglu & Restrepo
Journal of Political Economy, 2020
The empirical counterpart to the task model: one more industrial robot per thousand workers reduces local employment and wages, concentrated in manufacturing communities. Displacement is not hypothetical — it shows up in commuting-zone data.
The benchmark causal estimate of automation's labor-market bite, and the empirical anchor of the AI-and-labor debate.
2021
EmpiricalPolicy
The Short-Run Effects of the General Data Protection Regulation on Technology Venture Investment
Jia, Jin & Wagman
Marketing Science, 2021
The best early evidence on what DATA regulation does to innovation: after GDPR, venture investment in EU technology firms fell relative to the US — fewer deals and smaller amounts, concentrated in data-intensive and younger ventures.
The empirical benchmark in the AI-regulation debate: privacy rules have real, measurable costs on the startup margin — to be weighed against their benefits.
2022
EmpiricalAutomation & Labor
Artificial Intelligence and Jobs: Evidence from Online Vacancies
Acemoglu, Autor, Hazell & Restrepo
Journal of Labor Economics, 2022
AI adoption measured from the universe of online job postings: establishments whose task mix is AI-suitable post more AI jobs and FEWER of the jobs AI can do — direct vacancy-level evidence of task displacement, though aggregate employment effects remain hard to detect.
The microdata bridge between the task theory and the labor market — AI's fingerprints found in what firms stop hiring for.
AnalyticalPolicy
Generative AI for Economic Research: Use Cases and Implications for Economists
Korinek
Journal of Economic Literature, 2023
A living field manual: LLMs as research assistants across ideation, writing, coding, data work, and math — with honest benchmarking of what they do well, where they hallucinate, and how economists' comparative advantage shifts as capabilities grow.
The profession's reference document for AI-assisted research — updated as the models improve, which is itself the message.
2023
AnalyticalPolicy
Regulating Transformative Technologies
Acemoglu & Lensman
American Economic Review: Insights, 2024
A theory of OPTIMAL adoption speed for technologies that are transformative but carry hard-to-reverse risks: when damages are irreversible and learning happens with use, the optimal policy is gradual adoption that accelerates as uncertainty resolves — not bans, not laissez-faire.
The formal backbone for 'go slow, then fast' AI governance — regulation as sequencing, not prohibition.
AnalyticalData Economics
The Changing Economics of Knowledge Production
Abis & Veldkamp
Review of Financial Studies, 2024
Knowledge work has a production function — labor plus data — and machine learning changes its shape: data becomes substitutable for analyst labor. Estimated on the finance industry, the model quantifies how AI reallocates value from human capital to data owners.
The framework for the deepest labor question in finance: who captures the rents when the analyst's knowledge is in the training set?
2024
EmpiricalData Economics
Artificial Intelligence, Firm Growth, and Product Innovation
Babina, Fedyk, He & Hodson
Journal of Financial Economics, 2024
Measured AI adoption from millions of job postings and resumes: AI-investing firms GROW — more sales, more employment, more market share — and the mechanism is product INNOVATION, not cost-cutting. Growth concentrates among the largest firms, deepening industry concentration.
The first comprehensive firm-level evidence on what AI investment actually does — growth through new products, with winner-take-most dynamics.
AnalyticalGrowth
The Simple Macroeconomics of AI
Acemoglu
Economic Policy, 2025
The cold-shower paper: mapping task-level AI capabilities into aggregate productivity implies modest gains — roughly a percentage point of TFP over a decade, not a growth revolution. Hype meets national accounts.
The rigorous pessimist's benchmark in the AI-growth debate — the number every optimist now has to argue against.
2025
2020s
2023
EmpiricalProductivity
Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence
Noy & Zhang
Science, 2023
The first prominent RCT of ChatGPT at work: mid-level professional writing tasks completed 40% FASTER with 18% higher quality — and the productivity gains are LARGEST for the weakest writers. AI compresses the skill distribution.
The equalization result: generative AI lifts the bottom more than the top — with deep implications for wage inequality and skill formation.
EmpiricalProductivity
Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality
758 BCG consultants, randomized: on tasks INSIDE the AI frontier, consultants with GPT-4 were 12% faster and 40% higher quality; on tasks OUTSIDE it, AI users did WORSE — trusting the machine where it fails. Introduced 'centaurs' and 'cyborgs' as working styles.
The 'jagged frontier' is now the standard mental model for AI at work: the technology's capability boundary is invisible, and knowing where it lies is the new skill.
EmpiricalCapital Markets
Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models
Lopez-Lira & Tang
Working paper (University of Florida), 2023
Asked ChatGPT to score news headlines as good or bad for the stock — no finance training, no fine-tuning. The scores predicted next-day returns, outperforming traditional sentiment measures, especially among smaller stocks. The market had not yet priced the machine's reading.
The paper that launched a thousand LLM-trading studies — and an early exhibit on how fast such predictability decays once everyone has the same machine.
EmpiricalCapital Markets
Can ChatGPT Decipher Fedspeak?
Hansen & Kazinnik
Working paper (Federal Reserve Bank of Richmond), 2023
GPT models classify the hawkishness of Fed statements about as well as trained human readers — and better than dictionaries and smaller models — while explaining their reasoning. Central-bank communication, machine-read.
Evidence that LLMs match domain experts on nuanced institutional text — from inside the Federal Reserve System itself.
EmpiricalAgents & Simulation
Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?
Horton
NBER Working Paper No. 31122, 2023
Put LLMs INSIDE the experiments: GPT agents endowed with preferences replay classic behavioral economics studies — Kahneman-Knetsch-Thaler fairness, status quo bias, minimum-wage hiring — and broadly reproduce human results, cheaply and at scale.
Founded the 'homo silicus' agenda: LLM agents as wind tunnels for piloting experiments and simulating populations before spending on human subjects.
EmpiricalDisclosure
Bloated Disclosures: Can ChatGPT Help Investors Process Information?
Kim, Muhn & Nikolaev
Working paper (University of Chicago), 2023
GPT summaries of earnings calls and MD&As are dramatically SHORTER yet capture the price-relevant content — and the 'bloat' the machine strips out is itself informative: bloated disclosures are associated with worse information environments.
Turned the LLM into a measurement device for disclosure quality: what the machine can delete without losing information is, by definition, noise.
EmpiricalCapital Markets
Generative AI and Firm Values
Eisfeldt, Schubert & Zhang
Working paper (NBER No. 31222), 2023
Built firm-level exposure to generative AI from the task content of occupations — and showed the stock market repricing it within DAYS of ChatGPT's release. Firms whose workforces are more exposed to LLMs gained value; Artificial Minus Human portfolios earned significant returns.
The event study of the ChatGPT moment: markets capitalized the GenAI shock immediately, and unevenly across firms.
2024
EmpiricalCapital Markets
Financial Statement Analysis with Large Language Models
Kim, Muhn & Nikolaev
Working paper (University of Chicago Booth), 2024
Gave GPT-4 standardized, ANONYMIZED financial statements — no text, no names, no dates — and chain-of-thought prompts mimicking analyst reasoning. The model predicts earnings direction better than human analysts and matches specialized ML models, with narrative rationales investors can audit.
The provocative result for the profession: the reasoning layer of fundamental analysis, not just the reading layer, is now automatable.
EmpiricalProductivity
GPTs are GPTs: Labor Market Impact Potential of LLMs
Eloundou, Manning, Mishkin & Rock
Science, 2024
Task-by-task exposure mapping across the entire US labor market: about 80% of workers have at least 10% of their tasks exposed to LLMs, and roughly a FIFTH have half or more — with exposure INCREASING in wages and education. White-collar work is the frontier this time.
The exposure atlas the policy debate runs on: generative AI as a general-purpose technology, formally argued in the title's double meaning.
EmpiricalCapital Markets
ChatGPT and Corporate Policies
Jha, Qian, Weber & Yang
Working paper (NBER No. 32161), 2024
Asked GPT to read earnings-call discussions and score firms' expected INVESTMENT policy. The resulting 'GPT investment score' predicts actual future capex and R&D — and subsumes traditional text measures. LLMs extract managerial intentions, not just tone.
From sentiment to substance: LLMs recovering forward-looking corporate policy from the words executives choose.
EmpiricalBias & Ethics
More Human than Human: Measuring ChatGPT Political Bias
Motoki, Pinho Neto & Rodrigues
Public Choice, 2024
A clean design for measuring ideology in a chatbot: ask ChatGPT to impersonate partisans, then compare its DEFAULT answers to the impersonations. The default sits systematically left of center across countries — a measurable political lean in a tool used by hundreds of millions.
One of the most-cited AI-bias papers in social science — and the methodological template of asking the model to reveal its own priors.
2025
EmpiricalProductivity
Generative AI at Work
Brynjolfsson, Li & Raymond
Quarterly Journal of Economics, 2025
5,000+ customer-support agents at a real firm, staggered AI rollout: 15% average productivity gain, concentrated among NOVICE workers — the assistant embodies the tacit knowledge of top performers and transfers it down. Turnover falls; customers are happier.
The first major QJE evidence on GenAI in production: AI as a machine for diffusing expertise — with the returns flowing to the least experienced.
EmpiricalDisclosure
ChatGPT for Textual Analysis? How to Use Generative LLMs in Accounting Research
de Kok
Management Science, 2025
The methods paper for the LLM era in accounting: when generative models beat traditional NLP, how to prompt, validate, and audit them, and the reproducibility traps — model drift, temperature, versioning — that can silently break a study.
The field's instruction manual for doing LLM-based research that replicates.
2026
EmpiricalBias & Ethics
Impartial Intelligence? Country Bias in GenAI Financial Analysis
Pinto & Motoki
Journal of Business Ethics, 2026
Give GenAI the SAME financial fundamentals and change only the country: the analysis changes. Systematic country bias in AI-generated financial assessments — with direct consequences for cross-border capital allocation as investors outsource analysis to machines.
Bias audits arriving in finance: the training data's geography becomes the analyst's prejudice.
EmpiricalLearning
AI Adoption and the Loss of Learning
Pinto, Radhakrishnan & Weber
Working paper, 2026
The productivity gains of AI adoption have a shadow price: the on-the-job learning that junior work used to produce. If AI does the apprentice tasks, where do the next experts come from? Examines how AI adoption reshapes skill development inside firms.
The human-capital question underneath every AI productivity study — connecting the GenAI literature to the economics of learning-by-doing.