Can AI predict the stock market? What the research actually measures

Different numbers answer different questions

Earnings direction outside the model's training window

58.96%

Kim, Muhn and Nikolaev, Table 6 Panel C; not price direction or Indian-market trading accuracy.

Source: [2]

Nifty afternoon direction, as reported

62.9%

Zerodha's dated short experiment; no independent reproduction or future-performance claim.

Source: [5]

Individual derivatives traders losing money

87.7%

SEBI FY26 study; not a measure of AI forecasting accuracy.

Source: [6]

What would count as evidence for a stock forecast?

First identify the target: an opening price, a return over a specified period, an earnings change or a probability of a particular event. Then identify the information available when the answer was produced. A prediction recorded after most of the relevant news is known answers a different question from one recorded earlier.

For a claim about tradable results, ask how a position would be entered and closed, what costs apply and how losses are handled. Directional accuracy alone leaves these questions open. A report can correctly classify more outcomes than a simple reference rule and still fail to establish a usable result under another execution assumption.

The table is an evidence map, not a ranking of models. It records the outcome each source measured, its testing boundary and what it says about costs. 'Not established here' means the source review does not support that claim; it does not mean the authors necessarily ignored the subject everywhere. We did not download proprietary datasets, call a prediction API or reproduce these papers.

Read the target, sample and cost assumption together

Selected published sources reopened 14 October 2026. News paper version 3 and earnings paper version 2 linked; other cited PDFs read as retrieved. No original testing or pooled accuracy score.

Study or report	Claim and metric	Testing boundary	Costs and transfer limit
Lopez-Lira and Tang, arXiv version 3	Authors report headline-sentiment scores associated with subsequent US daily returns; a portfolio/statistical study rather than Nifty direction accuracy.	October 2021–December 2022; selected version places this beyond its stated model training cutoff.	Figure 2 compares 0, 5, 10 and 25 basis points per transaction with assumed close/next-open fills. No current Indian retail fee model established.
Kim, Muhn and Nikolaev, arXiv version 2	Future earnings direction. Table 6 Panel C: 58.96% accuracy and 63.91% F1 score, which combines precision and recall; these are different metrics.	Anonymous statements; separate 2022-input/2023-earnings analysis outside training window, with legacy-model replication discussed.	Constructed portfolio section includes 10-basis-point sensitivity. An earnings classifier result is not an Indian option-chain forecast.
Mehtab, Sen and Dutta, Nifty study	Opening-value regression evaluated using correlation, root mean squared error (RMSE) and related ratios; not a percentage of profitable trades.	Regression training 29 December 2014–28 December 2018; test 31 December 2018–31 July 2020. LSTM uses walk-forward validation.	A current India after-cost trading result is not established from these regression metrics. Historical price fit cannot be read as trade win rate.
Chen, Green, Gulen and Zhou	Authors report excessive extrapolation from recent returns, optimistic forecasts and interval-calibration issues.	Study of model forecasts versus historical/realised returns and crowd estimates; not a reproduced Nifty strategy.	No Indian net-trading-cost comparison established by this review. Calibration and prediction profitability are different questions.
In The Money by Zerodha, 2 June 2026	Reports 62.9% afternoon direction accuracy for Claude Opus 4.6 across its stated 31-day Nifty experiment.	24 March–12 May 2026; afternoon and morning inputs differ. Report itself warns the sample is small.	Hypothetical spot exercise changes under a later exit assumption; source discusses additional costs and slippage. No audited after-cost result reproduced here.

Sources: [1], [2], [3], [4], [5]

Download this table (CSV)

Out-of-sample is a question about information, too

In an ordinary model experiment, the test observations are separated from the observations used to fit the model. With a language model, an additional question arises: could the outcome or related discussion have been present in the model's earlier training material? Merely hiding a future price column from your spreadsheet does not answer that question.

The earnings paper explicitly addresses this issue through anonymised statements and a separate period beyond the model's training window. The news paper's selected version explains its own training-cutoff rationale. Those details belong to the stated research designs; they should not be generalised to every new chatbot or every historical chart prompt.

A fair assessment also retains failed forecasts and choices made while testing. If a model, prompt or horizon was selected after examining outcomes, that selection needs to be disclosed. Compare with a stated simple reference method and preserve the period that was not used to tune the approach. Our backtest and live-results guide explains the separate execution and data gaps without treating a chart fit as a trading recommendation.

How a near-100% price score can get direction wrong

Entirely hypothetical index levels, written for arithmetic education. These are not Nifty observations, model outputs or a backtest. The source IDs explain why regression and directional metrics must remain separate.

Assumed starting levelAn invented reference point.

20,000

Assumed forecast and realised levelForecast is up 10; the hypothetical outcome is down 10.

20,010 forecast; 19,990 realised

Absolute forecast error

|20,010 − 19,990| = 20

Absolute percentage errorRounded to five decimal places after calculating.

20 ÷ 19,990 × 100 = 0.10005%

A misleading 'accuracy' conversionUses the unrounded error, then rounds to five decimal places. This number is not a verified directional accuracy rate.

100 − (20 ÷ 19,990 × 100) = 99.89995%

Sources: [3], [2]

Ask what '99% accuracy' actually divides by

The hypothetical forecast sits close to the level but points in the opposite direction. That is the reason to ask for the metric definition before accepting a percentage. An error relative to a large price level, a correlation between two series and a count of correct directional classifications are different quantities.

For a directional classification, you need the number of correct calls and the number of eligible outcomes, along with the treatment of ties, missing responses and abstentions. For a price error, you need the forecast horizon, units and error formula. A probability estimate needs calibration: do events assigned a particular probability occur at an appropriate frequency across a suitable sample?

The extrapolation paper reports that models can repeat a familiar human bias by placing too much weight on recent performance. This is a published research finding in that study, not a claim that every model makes the same error on every input. It provides a reason to inspect the evidence behind an answer's confidence rather than interpreting confident prose as a calibrated probability. Return-forecast study.

Do not turn a model's self-reported confidence into a performance statistic. Ask whether it was measured against outcomes, on which sample and with what exclusions. If a seller supplies only an impressive percentage and no denominator or retained prediction record, the evidence needed to assess the claim remains missing.

Why the dated Nifty experiment needs its caveats

Zerodha's report separates the afternoon information set from the following morning's overnight context and says the sample is small. It also describes weaker hypothetical results when the assumed exit is delayed. These details are central to interpreting the published report, rather than footnotes to an accuracy headline. The model name is part of that historical experiment, not a recommendation about the current app. Zerodha's original report.

We quote its reported percentage, not an independently verified count of successful calls. Its approximate count does not allow the percentage to be reconstructed exactly from the stated sample, so this page does not derive additional wins, confidence intervals or significance from it. The raw prediction and execution records would be needed for a reproduction.

For any similar result, ask whether information arrived before the proposed decision time, whether a simple baseline was evaluated and whether a full cost model belongs to the result. A retrospective selection of a model is also different from a choice fixed before the test. Keep each uncertainty visible instead of treating a short demonstration as a general law about market predictability.

Questions for an AI forecast or paid trading coach

• What is the exact outcome?

  Request the instrument, horizon and metric definition. A claim about earnings or price fit should not be relabelled as daily trading accuracy.

• Where is the complete dated record?

  Look for predictions saved before outcomes, including incorrect calls, no-answer cases and later corrections. Selected screenshots cannot establish the denominator.

• What assumptions create the result?

  Identify information timing, data revisions, execution price, costs and model selection. Ask which assumptions were fixed before evaluating the sample.

• What does the claim promise?

  A guaranteed result is not supplied by the research cited here. Verify any required intermediary registration through the regulator rather than relying on a chatbot or promotional badge.

A 60-entry review can make a narrower, checkable statement

Illustrative invented journal categories; not real user trades, a chatbot output or evidence that journaling improves profits.

Assumed sampleEach contains enough notes to apply the stated rules.

60 historical entries

Assumed flagsCategories may overlap; they are not added blindly.

18 late entries; 12 changed stops; 4 in both

Distinct entries with either flag

18 + 12 − 4 = 26

Share of the sample flaggedRounded to two decimal places. A descriptive record count, not prediction accuracy or an expected future loss rate.

26 ÷ 60 × 100 = 43.33%

Source: [7]

Retrospective review needs evidence of its own

In the hypothetical journal, an assistant could point to the entries carrying each flag and ask whether its reading matches your notes. The original records give you something to inspect. The percentage does not prove why a trade lost money, whether a changed stop caused that loss or what will happen next time. It also cannot establish a benefit from using AI: that would require separate evidence and an appropriate study design.

Treat missing reasons as missing. A blank checklist is not proof that the trader ignored a rule, and a losing trade is not automatically a rule break. Ask the assistant to separate explicit notes, calculations and interpretation. A useful review can identify an unresolved question without prescribing a trade.

OneTradeJournal records trades manually, with no import or broker sync. Its paid AI can be wrong; it does not predict prices, provide signals, connect to a broker or block an order. Read about AI journal review for that narrower use and the model documentation comparison for file and source limits.

SEBI's PR 50/2026 reports that 87.7% of individual equity-derivatives traders lost money in FY26. The study is market-risk context, not a comparison of AI users against non-users and not proof that a particular tool caused or prevented losses. SEBI's release.

SEBI PR 48/2026 of 17 August cites circular HO/47/17/12(11)2025-MRD-POD3/I/11107/2026, dated 8 May 2026, for the thirty-day educational price-data lag. Its wording says a person engaged solely in education should not indicate future prices or provide securities recommendations. This is a plain-English summary, not legal advice; the examples above use no actual price data. SEBI's caution.

If you arrived looking for an AI Nifty prediction for tomorrow, there is no forecast here. The available evidence helps you ask what a prediction claim measured and what it left unresolved. An assistant's next sentence is not a substitute for that record.

Sources

Every figure on this page comes from one of these. Each was checked on the date shown.

1. [1]Lopez-Lira and Tang, arXiv 2304.07619 version 3 · checked 14 Oct 2026
2. [2]Kim, Muhn and Nikolaev, arXiv 2407.17866 version 2 · checked 14 Oct 2026
3. [3]Mehtab, Sen and Dutta, Nifty regression and LSTM paper · checked 14 Oct 2026
4. [4]Chen, Green, Gulen and Zhou, extrapolation and miscalibration · checked 14 Oct 2026
5. [5]Sandeep Rao, In The Money by Zerodha, 2 June 2026 · checked 14 Oct 2026
6. [6]SEBI PR 50/2026: FY25–FY26 derivatives studies, 20 August 2026 · checked 14 Oct 2026
7. [7]SEBI PR 48/2026: educational price data, 17 August 2026 · checked 14 Oct 2026