The opposing informational effects of AI trading
Organizations: Penn State, Great Valley School of Graduate Professional Studies
Abstract
Better artificial intelligence (AI) can improve individual forecasts without making prices better at revealing new information. In a competitive market with resource-constrained forecasting, more capable algorithms can substitute analysis of public records for private observations, reducing the private information aggregated into prices. The same trades can convey sophisticated public-data analysis to investors unable to perform it themselves. We establish conditions under which these effects coexist and the private benefit of further computation falls. Numerical equilibrium calculations and finite-data neural-market experiments exhibit corresponding patterns. Out-of-sample earnings forecasts and controlled language-model tasks show related patterns of information substitution. In the neural experiment, wider adoption increases the predictive value of price for limited observers, reduces it for advanced observers, and lowers the benefit of upgrading by 44%. Retraining by investors who did not upgrade can amplify the loss of private-information precision. The informational consequences of AI trading depend on what prices reveal and who can use them.
Figures & tables
| Object | Economic role | Who observes or chooses it? |
|---|---|---|
| , | Latent economy and the scalar target of private clues. | Latent primitives, inferred from private and public observations. |
| Account ’s noisy clue about . | Privately observed by account . | |
| Algorithm ’s processed summary of past public observations. | Computed from data accessible to everyone; the summary depends on capacity. | |
| Desired number of risky shares at candidate price . | Chosen by investor using , and . | |
| , | Clearing price and high-capacity adoption fraction. | emerges from clearing; specifies a counterfactual technology mix. |
| Stage | Events and information available |
|---|---|
| 0. State and releases | is realized; previously traded claim pays ; the public sees , , and the prior price , updating . The new payoff is not known. |
| 1. Pre-auction computation | Every deployed forecasting algorithm updates its internal state from . This step cannot use or . |
| 2. Private clues | Individual investors observe and choose their full price-contingent schedules . |
| 3. Clearing | The exogenous net supply is presented to the auction; is the price at which integrated demand equals supply. |
| 4. Public price | The newly determined enters . Forecasting rules are not refitted within this auction. |
| 5. Next date | matures and is disclosed before the next auction, entering . The process repeats with the next one-period claim. |
| Result | Informational and statistical specialization | Claimed comparison |
|---|---|---|
| Theorem 1 (S4) | Aligned payoff; optimally fitted constant-affine Gaussian head; arbitrary feasible nonlinear public encoders with nested complete algorithm classes; fixed inherited law. | Upgrade changes the aggregate private-signal loading and fresh-price likelihood precision in the first deployed auction, with other policies fixed. |
| Theorem 2 (S5) | Correct joint Gaussianity; aligned payoff; the same explicitly nested finite-capacity public information fields, with current price and own private signal added to both. | Information accounting and comparative-statics conditions for a negligible investor’s private capacity premium along stationary laws . |
| Theorem 3 (S6) | A two-factor, publicly released macro signal and a private fundamental innovation; population Gaussian forecasts and CARA competitive clearing; aligned payoff. | Strict local coexistence of negative new-information discovery, positive low-capacity price information, and lower private capacity premium, with both frozen policies and population refitting. |
| Label | Maintained in core economic environment | Added, rather than presumed, later |
|---|---|---|
| ENV-1 | Finite-dimensional stable exogenous economic state, stationary baseline covariance. | Reachability/observability or full-rank projections where required. |
| ENV-2 | One-period claim and risk-free account; payoff noise . | A long-lived resale-price recursion or state-dependent structural payoff volatility in robustness studies. |
| ENV-3 | Continuum of atomistic traders, Gaussian private signals and exact within-type aggregation. | Alternative cross-sectional signal structures, if used in a separate experiment. |
| ENV-4 | Independent noisy net supply, , and unconstrained positions. | Positive clearing denominator (downward-sloping aggregate demand) and invariant market-price law (S3). |
| ENV-5 | Common raw pre-price public record with only matured payouts; capacity-specific processed summaries. | Global existence of an optimal learned forecast and nested information fields for a particular theorem (S2, S5). |
| ENV-6 | CARA portfolio choice against an individually fitted forecast and competitive price taking. | Subjective Gaussian head for S3 closed form; correct joint Gaussianity for S5–S6. |
| Symbol | Definition |
|---|---|
| Fast trading date and slow deployment/retraining epoch, respectively. | |
| Latent state, transition, innovation loading, stationary state covariance. | |
| Dimension of the latent state (not computational capacity). | |
| Public macro-release vector, observation loading, and optional release noise. The scalar in S6 is a public macro factor. | |
| Privately signaled scalar fundamental and its latent-state loading. | |
| One-period payoff, intercept, factor loading, independent payoff shock. |
| Object | Meaning | Available when? |
|---|---|---|
| Economic state and the common fundamental; generally latent, not supplied to an investor. | Not directly observed. | |
| Investor-specific noisy observation of ; its realization is private. | Before submitting demand. | |
| Released macro data, past prices, and matured payouts; common raw record. | Before the auction. | |
| The algorithm’s computed representation of that record. | Computed before the auction. | |
| Candidate current clearing price in a demand schedule. | Varied at the auction. | |
| Target payout for forecasting and for the subsequent payoff realization. | Revealed only next period. |
| Implementation | How public history is represented | Resources to measure |
|---|---|---|
| Finite impulse response / linear state space | A fixed lag window or a linear recurrent state; includes the analytically solvable Gaussian benchmark. | State memory, coefficients, matrix operations. |
| GRU / recurrent network | Nonlinear recurrent update of a learned hidden representation. | Weights, stored hidden state, update/forecast operations. |
| Finite-context transformer | History processed by attention over a bounded token/window sequence or retained cache. | Context/cache memory, parameter storage, attention/readout operations. |
| Finite-history feature program | A prescribed collection of lags, thresholds, and nonlinear transformations of public data. | Stored history, instructions, output/head costs. |
| Model A: constant affine head | Model B: state-dependent affine head | |
|---|---|---|
| Predicted mean | ||
| Predicted variance | One fitted number per type | Fitted function |
| Public-data computation | Arbitrary admissible pre-auction nonlinear encoder | Arbitrary admissible pre-auction nonlinear encoder |
| Principal role | General first-deployment sign theorem (S4) | Richer learning and risk experiments (S8–S9) |
| Result | Training and head restrictions | Informational comparison |
|---|---|---|
| Theorem 1 (S4) | Model A; aligned payoff; complete nested feasible algorithm classes; free optimal constant affine readout and variance; attained global population minima. | Same inherited and its historical ; a first-auction price-signal precision comparison. |
| Theorem 2 (S5) | Correct joint Gaussian market law; candidate low and high conditional forecasts are both fitted correctly. | Same , current price and private signal for both, and explicitly nested feasible finite-capacity public information fields. |
| Theorem 3 (S6) | Linear-Gaussian two-factor specialization; Gaussian-optimal forecasts; CARA-optimal holdings and clearing. | Locally endogenous adoption path, frozen incumbents or all-type population refitting; signs proved near zero adoption. |
| Symbol | Definition |
|---|---|
| Learned forecasting system, encoder parameters, public-state update, resulting pre-auction state. | |
| Inference-resource budget, memory and operation costs, feasible full forecasting systems. | |
| Fitted payout mean and variance evaluated at a candidate current price. | |
| Model-A affine head intercept, public-state loading, private-signal loading, price loading. | |
| Model-B state-dependent mean readout functions. | |
| Inherited/induced stationary training law, type- marginal, realized current price in training samples. |
| Stage | What happens | Information available |
|---|---|---|
| Before auction | Public reports, and matured payout arrive; the AI constructs ; investor observes . | and own . |
| Orders and clearing | The trained head evaluates each candidate ; the investor submits ; all schedules clear at . | Candidate prices in schedules; no individual takes as personally controllable. |
| After clearing | The realized becomes public. At the next date, is paid and revealed. | . |
| Slow learning epoch | A sequence of trades generates training observations ; population or sample learning may refit the forecasts. | Realized historical prices and matured payout labels. |
| SI proof | S3 input | What must still be established there |
|---|---|---|
| S4 / Theorem 1 | Demand slope ; inherited-law ; likelihood precision . | Nested optimal learning, aligned payoff and the correct sign of private-signal response. |
| S5 / Theorem 2 | Common price and pre-price low/high information fields; stationarity and, where used, jointly Gaussian law. | Information identity, Gaussian posterior variances, ex-ante capacity value and its derivative. |
| S6 / Theorem 3 | Price formed by optimal CARA demands and actual clearing; locally differentiable equilibrium equations. | A two-factor primitive economy with strict coexistence signs and an open neighborhood. |
| Symbol | Meaning |
|---|---|
| Individual CARA-optimal holding at candidate clearing price . | |
| Learned subjective Gaussian conditional mean and state-dependent predictive variance. | |
| Population mass divided by risk aversion and fitted predictive variance. | |
| Aggregate public-history component of demand, before subtracting the supply level. | |
| Aggregate loading on the privately signaled scalar fundamental . | |
| Positive aggregate price slope magnitude; is the aggregate demand derivative. |
| Low capacity | High capacity | |
|---|---|---|
| Projection residual | ||
| Private-signal forecast coefficient | ||
| Fitted payoff variance | ||
| Private-signal demand response |
| Initial values | Changed values | |
|---|---|---|
| Low-capacity uncertainty after price, | ||
| High-capacity uncertainty after price, | ||
| Price information after limited analysis, (nats) | ||
| Price information after advanced analysis, (nats) | ||
| Gross upgrade premium, (payoff units) |
| Baseline | Retrained derivative | Frozen derivative | |
|---|---|---|---|
| Weight on in , | |||
| Weight on in , | |||
| Price information given , (nats) | |||
| Price information without , (nats) | |||
| Upgrade premium, (payoff units) |
| Quantity | Value or construction |
|---|---|
| Latent state; largest transition eigenvalue modulus | 8 factors; |
| Public macro releases | 2 noisy indicators; noise SDs |
| Payout coefficient and independent shock SD | , |
| Net supply shock SD; risk-free gross return | , |
| Group masses | |
| Risk aversion |
| Before adoption | First deployment | Retrained equilibrium | |
|---|---|---|---|
| High-capacity share within group 1 | 0 | 1 | 1 |
| Fresh-price precision | 2.469507 | 2.262529 | 2.073514 |
| Change relative to baseline | — | ||
| Local retraining Jacobian radius | 0.6871 | — | 0.7236 |
| Fast-market transition radius | 0.9400 | — | 0.9400 |
| , frozen | , retrained | ||
|---|---|---|---|
| , prior variance after | 0.422095 | 0.422095 | 0.422095 |
| , prior variance after 3 lags | 0.173595 | 0.173802 | 0.174489 |
| , posterior variance after price (limited) | 0.206670 | 0.142132 | 0.160367 |
| , posterior variance after price (high) | 0.121506 | 0.124747 | 0.128131 |
| Price information (nats) | 0.357055 | 0.544236 | 0.483883 |
| Fresh price information (nats) | 0.178380 | 0.165813 | 0.154406 |
| Forecasting head deployed for upgrading group | Change from original | |
|---|---|---|
| Old mean, old variance | 2.469507 | — |
| New mean, old variance | 2.108368 | |
| Old mean, new variance | 2.735404 | |
| New mean, new variance | 2.262529 |
| Groups permitted to refit their forecasts | Fresh-price |
|---|---|
| Neither (initial deployment, coefficients fixed) | 2.262529 |
| Only upgrading group 1 | 2.272432 |
| Only non-upgrading group 2 | 1.994346 |
| Both groups | 2.073514 |
| Condition satisfied | Frozen policies | Population retraining |
|---|---|---|
| Price information to limited observer increases | 147/147 | 147/147 |
| Fresh information beyond three public lags decreases | 147/147 | 147/147 |
| Marginal candidate’s computation value decreases | 147/147 | 147/147 |
| High-capacity pre-price variance nondecreases | 76/147 | 108/147 |
| Limited posterior variance falls and high posterior variance rises | 140/147 | 145/147 |
| Initial deployment | Endogenous-history and retraining experiment | |
|---|---|---|
| Public-history encoder | Linear state-space model (SSM) or gated recurrent unit (GRU) | GRU for group 1; no public-history encoder for group 2 |
| Model sizes | 4 and 12 hidden coordinates within each architecture | 4 and 32 hidden coordinates in group 1 |
| Historical window | 48 observations | 16 observations |
| Forecasting head | Constant affine Gaussian (Model A); selected Model-B comparisons | Constant affine Gaussian (Model A) |
| Primary comparison | Prediction loss and fresh-price precision at the first auction | Precision after endogenous price-history adjustment and up to 12 or 6 model updates |
| Independent learning runs | Three matched initial fitting seeds in the aligned comparison | Three inherited trained initial seeds, reused in paired adoption and retraining comparisons |
| Architecture | Hidden dimension | Held-out negative log score | Fresh-price |
|---|---|---|---|
| Linear SSM | 4 | 0.94043 | 0.27184 |
| Linear SSM | 12 | 0.93946 | 0.26559 |
| GRU | 4 | 0.95134 | 0.27605 |
| GRU | 12 | 0.94584 | 0.27392 |
| Object | Implemented value |
|---|---|
| Investor masses; risk aversion; private noise variance | ; ; |
| Group-1 public-data models | GRU widths 4 and 32; Model-A constant-affine mean and variance head; 16-date window |
| Group-2 model | Affine mean in private signal and current price; constant fitted variance; no public encoder |
| Adoption and initial random seeds | ; seeds , , |
| Training observations per update | 512 independent paths, 2,048 usable dates each; 1,048,576 observations |
| Validation and held-out evaluation | 65,536 and 131,072 observations, respectively; independent shocks |
| Group-1 adoption | End-point | High-model MSE gain | High versus low |
|---|---|---|---|
| 3.3489 | |||
| 3.1737 | |||
| 3.0647 |
| Contribution | Group-1 adoption 50% | 100% |
|---|---|---|
| Direct adoption, coefficients fixed | ||
| Additional retraining by group 1 alone | ||
| Additional retraining by group 2 alone | ||
| Additional interaction between the groups | ||
| Both-group additional retraining | ||
| Total: adoption and both-group retraining |
| Group-1 adoption and comparison | Mean channel | Variance channel | Net effect |
|---|---|---|---|
| 50%: initial adoption | |||
| 50%: both groups’ additional retraining | |||
| 100%: initial adoption | |||
| 100%: only group 2 refits, additional | |||
| 100%: both groups’ additional retraining |
| Item | Implemented choice |
|---|---|
| Market profiles | Three inherited seeds group-1 adoption ; both groups refitted for six updates |
| Candidate information | Width-4 or width-32 GRU of 16 pre-auction public observations; current price added at the forecasting stage; own private clue added for utility evaluation |
| Candidate learning | paths observations; Adam updates for each capacity; independent validation paths |
| Affine-head calibration | paths observations, separate from neural training and evaluation |
| Held-out evaluation | Four independent banks, each paths dates; burn-in dates per path |
| Statistical comparisons | Three paired initial seeds; descriptive Student- intervals ( degrees of freedom); additional path-level diagnostics |
| Adoption by group 1 | Smaller | Larger | Difference |
|---|---|---|---|
| Adoption by group 1 | Learned neural head | Recalibrated affine head | Both encoders and affine head |
|---|---|---|---|
| Experiment | What changes | What remains fixed |
|---|---|---|
| Training and prediction | Training length and random training samples for negligible candidates; supervised readout used to evaluate the price | Nine six-update markets (three seeds ), incumbent trading rules, and economic primitives |
| Market learning | More rounds of refitting by market participants; newly drawn versus reused primitive shocks | Three original market seeds, underlying economy, model classes, and per-update learning budget |
| Payout laws | Whether payouts are aligned with private clues and whether payoff noise varies with a public state | Market clearing, model classes, matched primitive draws, and a common first-deployment/frozen-history protocol |
| Candidate training | : low-minus-high price value | : upgrade value |
|---|---|---|
| Original training | ||
| One extra block | ||
| Three extra blocks | ||
| New sample, first refit | ||
| New sample, second refit |
| Forecast readout | Small GRU versus large GRU | Small GRU versus combined representations |
|---|---|---|
| Affine ridge | ||
| 64 random features | ||
| 128 random features |
| Learning samples / adoption | Change in at update 12 ( ) | Passes settling screen | Period-two flags |
|---|---|---|---|
| Fresh samples, | |||
| Fresh samples, | |||
| Fresh samples, | |||
| Reused shocks, |
| Fresh samples | Reused shocks | |||
|---|---|---|---|---|
| Forecasting rule | Forecast | Holding | Forecast | Holding |
| Small GRU | ||||
| Large GRU | ||||
| Other group | ||||
| Economy | Initial change ( ) | (dividend MSE) | (certainty equivalent) |
|---|---|---|---|
| Matched aligned control | |||
| Nonaligned payout | |||
| Public-state-dependent risk |
| Exercise | Computational procedure | Materials required |
|---|---|---|
| Candidate training | Hold the nine neural markets fixed; refit candidate predictors or change the readouts; calculate and on new evaluation paths | Saved market policies, training settings, candidate fits, evaluation samples |
| Market retraining | Continue both investor groups from update six; evaluate , the coefficients, and one-step policy changes | Market checkpoints, shock streams, learning configurations, seed-level statistics |
| Alternative payoffs | Refit the inherited laws for each payout specification; clear new adoption mixes; evaluate candidates on fixed-policy laws | Economic parameters, inherited model fits, paired shocks, evaluation summaries |
| Diagnostic | Count or share |
| Eligible CRSP monthly common-stock observations | 872,009 |
| Unique quarter-end PERMNO–forecast-date rows, 2011–2025 | 163,912 |
| Distinct PERMNOs at forecast dates | 7,807 |
| Annual-accounting match rate at forecast dates | 91.41% |
| I/B/E/S first future announcement within 180 days | 67.26% |
| Rows with a nonmissing future standardized EPS surprise | 58.26% |
| Model | Estimator | Inputs | Information available |
|---|---|---|---|
| Ridge | 14 | Limited accounting variables and lagged market data; no current snapshot | |
| Ridge | 17 | Exactly plus current log price, log market equity, and current one-month return | |
| Gradient boosting | 54 | All input families plus longer lagged return, price, volatility, and financial-statement histories | |
| Gradient boosting | 57 | Exactly plus the same three current-snapshot fields as |
| Contrast | Estimate | Two-way SE | Two-way 95% CI | Quarter-block 95% CI |
|---|---|---|---|---|
| Moment | Annual accounting | RDQ-timed quarterly accounting |
|---|---|---|
| Increment | (SE) | (SE) | ||
|---|---|---|---|---|
| Limited snapshot, | ||||
| Rich snapshot, | ||||
| Rich public, | ||||
| Rich after snapshot, |
| Increment | Coefficient (SE) | ||
|---|---|---|---|
| Sample or outcome | ||
|---|---|---|
| Annual, CAR | ||
| Annual, RDQ agreement within 1 day | ||
| Annual, CAR | ||
| Annual, BHAR | ||
| Quarterly, CAR | ||
| Quarterly, RDQ agreement within 1 day |
| OOS year or split | ||||
|---|---|---|---|---|
| 2018 | 2,333 | 0.000728 | -0.002317 | 0.003046 |
| 2019 | 4,783 | 0.001106 | 0.001427 | -0.000321 |
| 2020 | 9,820 | 0.003786 | 0.002325 | 0.001461 |
| 2021 | 11,311 | 0.004592 | 0.001446 | 0.003146 |
| 2022 | 8,921 | 0.003237 | -0.001424 | 0.004661 |
| 2023 | 5,661 | 0.002938 | 0.001043 | 0.001895 |
| Pre-event outcome | coefficient (SE) | coefficient (SE) | |
|---|---|---|---|
| Full interval, average daily | 55,696 | ||
| First 20 trading days | 42,073 | ||
| Last 20 trading days | 42,071 |
| Formation-date FE | Firm + formation-date FE | ||||
|---|---|---|---|---|---|
| 0 | 55,771 | ||||
| 1 | 43,557 | ||||
| 2 | 34,671 | ||||
| 3 | 28,404 | ||||
| 4 | 22,657 | ||||
| Input | What it contains | How it is used |
|---|---|---|
| Company accounts | Publicly reported financial numbers from before the forecast | Included in every prompt |
| Past market returns | 52 weekly and 60 recent daily observations, plus a return-based index | Included in the historical-price condition only |
| Independent research note | Hypothetical next-earnings-surprise estimate of or ; stated error SD | Varied between paired prompts |
| Future earnings and identity | Realized next earnings surprise, company name and ticker, dates, actual dollar price | Withheld from every prompt |
| Public information | GPT-5.2 | GPT-6.1 Sol | Difference | 95% interval |
| Accounts only | 0.7050 | 0.6835 | ||
| Accounts and past returns | 0.7492 | 0.7072 | ||
| Change in the model difference, | ||||
Algorithmic Trading
Algorithmic trading uses computational models to generate signals, allocate portfolios, and execute orders across financial markets.
See all 167 papers