Can AI Predict the Stock Market? What AI Forecasting Research Shows
Not reliably. Research finds that AI models can pick up some short-lived signals, such as how markets react to news, but those edges are small, fade as more traders use the same tools, and rarely survive trading costs. No AI can predict markets in the sense people hope for.
Not reliably, and the hope is the expensive part
People want a model that says up or down before the move, often enough to live on. The research that this page is allowed to cite does not support that hope. Models can pick up some short-lived signals, especially how markets react to news. Those edges are small. They fade as more traders use the same tools. They rarely survive trading costs. No AI can predict markets. The sense people hope for, a durable call on prices, is the sense the answer rejects.
The CFTC press release of January 25, 2024, Release 8854-24, says AI technology cannot predict the future. The customer advisory SI Won't Turn Trading Bots into Money Machines says SI technology can't predict the future or sudden market changes, and that promises of unreasonably high or guaranteed returns are a red flag. A page that leads with a guaranteed hit rate is the red flag, not a counterexample.
Crypto prices get the same answer as stock prices. A faster tape does not make a call easier to live on. It makes the fee show up sooner, because you can trade more often. Solana volume bots add a further problem: some of the tape is manufactured. A model that treats volume as truth will treat a manufactured tape as truth. That is not a subtle failure. It is a wrong input.
What one abstract actually reports
The abstract of arXiv 2304.07619 says GPT-4 headline scores capture the initial market reaction, about 90% portfolio-day hit rates for the non-tradable initial reaction, and some later drift, especially for small stocks and negative news, and that strategy returns decline as LLM adoption rises. Hold the adjectives that are doing the work inside that sentence. Initial. Non-tradable. Some later drift. Decline as adoption rises.
Headline scores and a tradable order are different objects. The high hit rate is attached to a reaction the abstract itself calls non-tradable. By the time an order can be sent, you are in the "some later drift" clause, and you are in a world where more people using similar tools shrink the return. AI stock pickers uses the same abstract and then prices the cost of acting on a list. AI chart analysis is the image version of the same temptation: a clean label, a later fill.
Why a market resists a durable call
A price is already a crowd's call. If a pattern is obvious and tradable, orders arrive until the pattern is no longer worth the fee. That is not a slogan about efficiency. It is the adoption sentence in the abstract, said in plain clothes. The more the tool is shared, the less the leftover is. A private trick that stays private is a different story, and it is not a story you can buy from a chat window.
Costs finish the job the crowd started. A call that is right slightly more than half the time, on a small move, pays the fee and then pays it again on the way out. The table below is that arithmetic at a 1.5% round trip, which is 0.75% each way once the ticket is large enough that the minimum fee does not bind. Small tickets are worse than 1.5%. The table is already enough to retire several popular hit rates.
Does AI trading work and is AI trading legit split the neighboring questions: evidence, and pitches. This page stays on accuracy versus cost. AI stock trading and AI trading bot are the product pages if you still want a tool after you see the table. A tool that summarizes is intact. A tool sold as a caller of prices is the hope this page is declining.
Worked example: what 55% accuracy is worth
Assumed numbers for the math, not a quote and not a price call. Example, not a recommendation. Assume a round trip costs 1.5%. Assume the move, if you are right or wrong, is the same size either way. At 55% accuracy the expected result is 0.55 times the move, minus 0.45 times the move, minus the 1.5% cost. That simplifies to a tenth of the move, minus 1.5 points.
| Move, either way | Expectancy at 55% accuracy | Accuracy that would only break even |
|---|---|---|
| plus or minus 1% | -1.40% | 125% (impossible) |
| plus or minus 5% | -1.00% | 65.0% |
| plus or minus 10% | -0.50% | 57.5% |
Read the first row twice. A 1% move, caught with a 55% hit rate, expects to lose 1.40% per try after a 1.5% round trip. Breaking even would require 125% accuracy. That is impossible. You cannot be right more than always. The fee is larger than the edge a 55% caller has on a 1% move. No prompt fixes the row.
The 5% row still expects to lose 1.00% per try at 55% accuracy. You would need a 65.0% hit rate just to tie. The 10% row expects to lose 0.50% per try at 55%, and it ties around 57.5%. Those are high bars for a call on a noisy tape, and they assume you actually catch the full move, both ways, with no extra slippage. Swaps and agent market trades pay 0.75% with a $0.90 minimum per swap. Limit, take-profit, stop-loss and trailing orders pay a flat 0.75% with no minimum. Onchain fees apply too. A 0.75% swap fee (minimum $0.90 per swap), plus onchain fees. Fees on your agent's trades earn up to 10% cashback in USDC. It is not a discount on the swap fee. Price a real ticket on the risk reward calculator. If your move is smaller than your round trip, stop arguing about the model.
What is still worth doing
Use the model to read, to list what would invalidate a plan, and to check that your size fits a loss you can take. Those jobs do not need a hit rate. AI trading prompts collects drafts for that clerical work. A draft is not a call. If the draft starts offering prices for next week, delete that part and keep the checklist.
Turn on Ask before every trade and your AI asks you first. Moby does not hold or approve trades. You set your agent's per-trade and daily spending limits and market-cap range; Moby blocks red-rated tokens. Limits are only as tight as you set them. Sells and exits skip the market-cap and safety checks, and tokens with no safety rating are not blocked. One agent per user, with its own wallet. Agents trade Solana tokens only. The agent cannot withdraw, transfer or send funds. AI Trading and copy trading currently need an access pass in the app. Supported hosts: Codex, Claude Code (CLI), Gemini CLI (with a paid Gemini API key or enterprise access) and Grok CLI. ChatGPT: works on Business, Enterprise and Edu workspaces that allow custom MCP servers (web). Download Moby if you want the app. The agent can help you execute a plan you already accept, inside caps. It is not a machine that makes the first row of the table possible. Nothing makes 125% possible. If a seller's page needs the first row to look good, you already have the advisory's red flag.
The same table on a faster tape
Crypto does not get a friendlier version of the arithmetic. A memecoin can move plus or minus 10% in an afternoon, which is the third row, and it can also go to zero, which is worse than the symmetric loss the table assumes. The table is kind. It assumes the wrong call loses only the same percent the right call gains. A token that can gap through your exit breaks that kindness. Size for the gap, not for the 55% story.
Short-lived headline reactions, the kind the abstract describes, are even harder on a tape that reprices in minutes. The non-tradable initial reaction is over before a careful reader has checked the token address. What remains is the later drift, which the abstract already shrinks, plus the fee, which the table refuses to ignore. Checking the address is the job. Calling the next hour is the job to decline.
Frequently asked questions
Can AI predict the stock market?
Not reliably. Models can pick up some short-lived signals, such as the way prices react to news, but the edges are small, they fade as more people use the same tools, and they rarely survive costs. No AI can predict markets in the sense people hope for. A guaranteed hit rate is a red flag.
Can AI predict crypto prices?
The same answer. A faster tape does not make a durable call easier. It lets you pay the fee more often, and some of the tape can be manufactured. No AI can predict markets. Treat a crypto caller with a guaranteed return the way you would treat that claim on a stock.
What did the headline-score study find?
The abstract says GPT-4 headline scores capture the initial market reaction, about 90% portfolio-day hit rates for the non-tradable initial reaction, and some later drift, especially for small stocks and negative news. It also says strategy returns decline as LLM adoption rises.
Why do these edges fade?
A tradable pattern attracts orders until the leftover is smaller than the fee. The abstract says returns decline as LLM adoption rises. Sharing the tool is part of the fading. A hit rate on a reaction you cannot trade is not a hit rate you can take home.
How accurate does a direction call need to be?
With a 1.5% round trip, a plus or minus 1% move would need 125% accuracy to break even, which is impossible. A plus or minus 5% move needs 65.0%. A plus or minus 10% move needs 57.5%. At 55% accuracy all three rows still expect a loss. Those bars assume you catch the full move.
Does a high hit rate pay after costs?
Not by itself. At 55% accuracy the example expects -1.40%, -1.00%, and -0.50% on plus or minus 1%, 5%, and 10% moves. The fee can be larger than the edge. Raise the move or cut the number of tries. Do not raise your faith in the hit rate alone.
What is AI useful for if it cannot call prices?
Reading, checklists, and sizing a loss you can take. Those jobs do not need a hit rate. Turn on Ask before every trade and your AI asks you first. Moby does not hold or approve trades. Use the model to test a plan you wrote. Do not use it as a caller.