AI price prediction accuracy: what 55–65% means and why 90% claims fail
Careful peer-reviewed directional accuracy clusters at roughly 55 to 65%, and reported figures above 90% are widely attributed to look-ahead bias. Published anchors include 58.07% and 64.21%, with a 2025 transformer study reporting about 69.1%. The more important point is that accuracy is not performance: a 58% hit rate produces profit only if the average gain exceeds the average loss by enough to cover trading costs, and in most retail settings it does not.
Published 27 August 2026 · Updated 28 August 2026 · AI Trading Book Editorial · Reading time about 13 minutes
- The realistic band is 55–65% directional in careful work — a few points above a coin flip, not a solved problem.
- The baseline is not 50%. It is the best trivial predictor on the same data, which is usually the majority class.
- Accuracy ignores magnitude, cost and frequency — the three things that decide whether a signal makes money.
- 90%+ figures usually come from shuffled splits, which put future information into training data.
- Five questions expose almost any accuracy claim, and a vendor who cannot answer them has not reported a result.
The published figures
These are the anchors from peer-reviewed and academic work. They are not directly comparable to each other — different assets, horizons, periods and target definitions — so read them as a band rather than a ranking.
| Study | Reported accuracy | Notes |
|---|---|---|
| Yu and Yan | 58.07% | Directional accuracy anchor at the lower end of the careful range |
| Ding et al. | 64.21% | Directional accuracy anchor at the upper end |
| Transformer study, ScienceDirect 2025 | approximately 69.1% | Above the earlier range; different architecture, task and period |
| Sirignano and Cont, Quantitative Finance, 2019 | approximately 65–75% deep learning; 57–67% linear | High-frequency task across 500 NASDAQ stocks; short horizon |
| Claims of 80–90% and above | Not credible as reported | Attributed to look-ahead bias from shuffling before the train/test split — arXiv 2407.09831 (preprint) |
What directional accuracy actually measures
Directional accuracy is the proportion of times a model correctly predicted the sign of the next move over a defined horizon on a defined test set. That is all. It carries no information about how large the moves were, how expensive it was to act on each prediction, or whether the test period resembled the conditions in which the model would be used.
Three specification choices change the number substantially without changing the model. The horizon matters — predicting the next tick and predicting next month's close are different problems with different achievable accuracy. The target definition matters, since predicting a move larger than some threshold is easier or harder than predicting any move at all, and studies rarely use the same threshold. And the class balance in the test period matters, which brings us to the more serious issue.
The baseline is not 50%
Comparing an accuracy figure to a coin flip is intuitive and usually wrong. The correct reference is the best trivial predictor on the same test data, which for a binary target is normally the majority class.
Suppose a test period contained upward moves 56% of the time. A model that outputs "up" unconditionally, with no inputs and no training, scores 56%. Against that baseline, a reported 58% represents about two points of genuine information, not eight. And because equity markets have trended upward over most long historical samples, up-days are frequently the majority class — meaning the baseline in an equity study is typically above 50%, sometimes well above it in a bull period.
This is why a study that reports accuracy without reporting the class balance of its test set has not reported enough to interpret. The omission is common, and it flatters results in exactly one direction.
Why 58% accuracy does not mean profit
Even a genuine edge over the correct baseline does not translate into money. Profitability is determined by an expression that accuracy appears in but does not dominate: hit rate times average gain, minus miss rate times average loss, minus costs per trade, all multiplied by the number of trades.
Four mechanisms break the link between the two.
Payoff asymmetry
Accuracy weights every prediction equally; profit does not. A model right 58% of the time whose average winner is smaller than its average loser loses money, and this profile is common in practice — models that clip small gains reliably tend to be exposed to occasional large adverse moves. The reverse also holds: a 45% hit rate with large winners and small losers can be strongly profitable. Accuracy simply does not rank strategies.
Costs per trade
Every prediction acted on incurs a spread, usually a commission, and slippage between the decision price and the fill. On a short horizon these costs are frequently larger than the average predicted move. A signal with a genuine two-point edge over baseline, generating a small expected move per trade, can be entirely consumed by the cost of taking the position — the edge exists and is unharvestable.
Turnover
Costs scale with the number of trades while the edge per trade stays fixed, so a high-frequency signal must clear a much higher bar than a low-frequency one. This is the reason accuracy achieved on very short horizons, such as the high-frequency order-book task in Sirignano and Cont, does not imply a tradable retail strategy: at that horizon the cost per decision is enormous relative to the move being predicted, and capturing it requires infrastructure a retail participant does not have.
Capacity
An edge that survives costs may still not survive size. Acting on a signal moves the price against you, and the effect grows with position size relative to available liquidity. A strategy that works with a small account can stop working when scaled — which is one reason published research results and deployed fund performance diverge.
| Question | Does accuracy answer it? |
|---|---|
| How often was the direction right? | Yes — this is exactly what it measures |
| How large were the correct moves? | No |
| How large were the incorrect moves? | No |
| What did it cost to act on each prediction? | No |
| How many trades were generated? | No |
| Would the edge survive at size? | No |
| Did the model beat a trivial predictor? | Only if the class balance is also reported |
| Would it have worked in a different regime? | No |
Where the 90% figures come from
Reported accuracies of 80 to 90% and above are, in the great majority of cases, artefacts of information leaking from the future into the training process. A preprint on arXiv (2407.09831) attributes such figures specifically to look-ahead bias introduced by shuffling data before splitting it into training and test sets.
Why shuffling destroys a time-series test
Standard machine-learning practice shuffles a dataset before splitting it, because for independent observations that removes ordering artefacts. Financial time series are not independent: adjacent observations are highly correlated, and the order is the substance of the problem.
When time-ordered data is shuffled, moments from after the prediction point land in the training set. The model is then asked to predict a point whose immediate neighbours it has already memorised. That is interpolation between known values, not forecasting, and it produces very high test accuracy that collapses entirely on genuinely unseen future data. The failure is silent — nothing in the training process signals an error, and the metric looks excellent.
| Mechanism | How it happens | Effect |
|---|---|---|
| Shuffled split | Time-ordered data randomised before separating train and test | Largest single inflator; produces the 90%+ figures |
| Global scaling | Features normalised using means or ranges computed across the whole dataset | Leaks future distribution into past observations |
| Point-in-time errors | Fundamentals or revisions used at their reporting date rather than publication date | Model acts on information not yet public |
| Survivorship in the universe | Testing only on instruments that still exist today | Removes failures from the sample; flatters both accuracy and returns |
How the split should be done
Chronologically, always. Train on an earlier period, validate on the period immediately after it, and hold out a final block that is never touched until the strategy is finished. Walk-forward testing repeats that arrangement across rolling windows, so every evaluation happens on data later than everything the model saw. Our backtest overfitting page covers the protocol and the statistics for judging whether a surviving result is real.
Reading Sirignano and Cont carefully
The strongest careful figures in the literature come from Sirignano and Cont's 2019 work in Quantitative Finance: approximately 65 to 75% for deep learning models against 57 to 67% for linear models, on a high-frequency task across 500 NASDAQ stocks. Two features of that result are routinely dropped when it is cited.
First, the comparison is the finding. The interesting number is not 75% but the gap over the linear baseline — evidence that a deep architecture extracted structure a linear model could not, on that task. Second, the setting is high-frequency order-book prediction, where the horizon is extremely short. Accuracy at that horizon says nothing about predicting daily or weekly direction, and capturing it demands latency and infrastructure that place it outside retail reach entirely.
Citing "up to 75% accuracy for AI in trading" from this paper, without the horizon and the baseline, converts a careful microstructure result into a marketing claim about a different problem.
Five questions that expose any accuracy claim
- Over what asset, horizon and period? Accuracy without these is uninterpretable.
- Against what baseline? What was the class balance of the test set, and what would a majority-class predictor have scored?
- Was the split chronological? If the data was shuffled, the number is not a forecast result.
- Accuracy of what, exactly? Direction of any move, or of a move exceeding a threshold, or something else?
- What is the net result after costs? Accuracy is an input to performance, never a substitute for it.
A provider unable to answer all five has not given you a result you can act on. The questions are also usefully diagnostic in the other direction: a vendor who answers them precisely, including the baseline, is unusual and worth taking more seriously.
Where this sits in the wider evidence
Modest directional accuracy is consistent with the rest of the record. A peer-reviewed review of 27 machine-learning equity experiments (Buczynski, Cuzzolin and Sahakian, 2021) found no conclusive evidence of machine-learning funds delivering returns at scale, and published AI fund indices have underperformed passive benchmarks over fifteen years — covered on our evidence page.
That coherence is the point. If 55 to 65% directional accuracy after costs produced a reliable edge, it would show up in fund returns, and it has not. The measured accuracy and the measured performance tell the same story from different directions.
What we could not establish
- A like-for-like accuracy comparison across architectures. Published figures use different assets, horizons, target definitions and periods, so no ranking of model types is available. Data not found.
- Any audited live directional accuracy for a retail AI trading product. Vendor figures are self-reported and generally lack a stated baseline or split method. Data not found.
- The class balance of the test sets behind several of the anchors above. Where a study did not report it, the effective edge over baseline cannot be computed. Data not found.
Key takeaways
- 55–65% is the honest band for careful directional prediction, with a high-frequency exception reaching higher on a different problem.
- Measure against the majority class, not a coin flip. Two points of real edge often hide inside eight points of apparent one.
- Accuracy is not performance. Payoff asymmetry, costs, turnover and capacity all sit between a hit rate and a profit.
- A shuffled split is the signature of a broken test and the usual explanation for 90%+ claims.
- Context strips most citations of their meaning. "Up to 75%" from a microstructure paper is not a claim about tomorrow's close.
- The accuracy evidence and the returns evidence agree, which is why neither supports the marketing claim.
Frequently asked questions
How accurate are AI stock price predictions?
Careful peer-reviewed directional accuracy clusters at roughly 55 to 65%. Published anchors include 58.07% (Yu and Yan) and 64.21% (Ding et al.), with a 2025 transformer study reporting about 69.1%. Sirignano and Cont reported approximately 65 to 75% for deep learning against 57 to 67% for linear models on a high-frequency task across 500 NASDAQ stocks.
Why do some tools claim 90% accuracy?
Reported accuracies of 80 to 90% and above are widely attributed to look-ahead bias, particularly shuffling time-series data before splitting it into training and test sets. Shuffling places observations from after the prediction point into the training data, so the model is scored on information it should not have had.
Is 58% accuracy good?
It is above a coin flip and, in a research context, potentially meaningful. It is not by itself evidence of a profitable strategy, because accuracy ignores the size of gains and losses, the cost of each trade, and how often the model trades. A 58% hit rate with losses larger than gains loses money.
Does a higher hit rate mean higher profit?
No. Profit depends on hit rate multiplied by average gain, less miss rate multiplied by average loss, less costs. A 45% hit rate with large winners and small losers can outperform a 65% hit rate with the reverse profile. Accuracy alone is not a performance measure.
What does directional accuracy actually measure?
The proportion of times a model correctly predicted the sign of the next move over a defined horizon on a defined test set. It says nothing about magnitude, nothing about the cost of acting on the prediction, and nothing about whether the test set resembled live conditions.
What is the baseline a model should beat?
Not 50%. The right baseline is the best trivial predictor on the same data, which is usually the majority class. If a test period had upward moves 56% of the time, a model that always says up scores 56%, so a reported 58% represents two points of information, not eight.
What is look-ahead bias?
Using information in a model that would not have been available at the moment of the simulated decision. Common forms include shuffling time-ordered data before splitting, scaling features using statistics computed across the whole dataset, and using fundamentals dated before their actual publication.
Why does shuffling time-series data break a model test?
Because adjacent observations are highly correlated. Shuffling puts moments from after the prediction point into training data, so the model is effectively interpolating between known neighbours rather than forecasting. Accuracy rises sharply and the result does not transfer to live trading.
How should time-series data be split?
Chronologically, never randomly. Train on an earlier period, validate on the next, and hold out a final untouched period. Walk-forward testing repeats this across rolling windows so the model is always evaluated on data later than everything it saw.
Do transformers predict prices better than LSTMs?
On some tasks, marginally. A 2025 study reported about 69.1% directional accuracy for a transformer approach, above the 58 to 64% range in earlier work, but architectures are tested on different assets, horizons and periods. Comparing accuracy figures across studies is not comparing model quality.
Does high accuracy on high-frequency data transfer to daily trading?
Not automatically. Sirignano and Cont's stronger figures came from a high-frequency order-book task where the signal horizon is very short and trading costs and latency requirements are correspondingly severe. Accuracy achieved at microsecond horizons is not evidence about predicting tomorrow's close.
How do I evaluate an accuracy claim from a vendor?
Ask five things: over what period and asset, against what baseline, split chronologically or randomly, accuracy of what quantity, and net of what costs. A vendor unable to answer all five has not given you a result you can interpret.
Can any model predict prices reliably?
No model predicts all future price movements reliably. Financial time series are noisy and non-stationary, and relationships that held in one regime can disappear. A peer-reviewed review of 27 machine-learning equity experiments found no conclusive evidence of machine-learning funds delivering returns at scale.
Is prediction the most useful thing AI does in finance?
The evidence suggests not. Documented value concentrates in execution cost reduction, fraud detection, surveillance and research throughput — tasks with a measurable objective and a clear counterfactual — rather than in unconstrained price forecasting.
Compiled by AI Trading Book Editorial from peer-reviewed and academic sources. Accuracy figures from different studies are presented as a range rather than a ranking, because assets, horizons, target definitions and periods differ. The 58/42 illustration is built on a published anchor and is labelled as an illustration, not a study result. The arXiv item is identified as a preprint. Items we could not trace are marked "data not found". Published 27 August 2026; last updated 28 August 2026. Corrections are logged on the corrections page.
Sources
- Sirignano and Cont, Quantitative Finance, 2019 — approximately 65–75% for deep learning against 57–67% for linear models on a high-frequency task across 500 NASDAQ stocks.
- Yu and Yan — directional accuracy 58.07%.
- Ding et al. — directional accuracy 64.21%.
- Transformer-based deep learning study, ScienceDirect, 2025 — approximately 69.1% directional accuracy.
- arXiv 2407.09831 (preprint) — reported accuracies of 80–90% and above attributed to look-ahead bias from shuffling before the train/test split.
- Buczynski, Cuzzolin and Sahakian, International Journal of Data Science and Analytics 11(3), April 2021 — review of 27 machine-learning equity experiments; no conclusive evidence of returns at scale.
Informational research only. Nothing on this page is personal financial, legal, tax or investment advice, or a recommendation to trade any instrument. Reported research accuracy is not an indication of live trading results.