AI and trading

AI Trading in 2026: What AI Can Do, What It Can’t, and How to Use It Right

By Alex J. Wang, Market Research, SEPT Investing · Published Nov 15, 2026 · Last updated Nov 15, 2026

Use AI to think and to write your rules. Let the rules do the trading.

Every week there's a new "AI trading bot" that promises to find the next big move for you. AI is genuinely useful for traders — but not in the way most of those ads suggest. This page explains, in plain words, where today's AI (the chat-style "large language models" behind most AI assistants) falls short in trading, what it's actually good at, and what kind of data makes it far more useful.

The reality check: AI trading on its own

In 2026, a startup called Nof1 ran public, real-money contests called Alpha Arena. Eight leading AI models each got $10,000 to trade US tech stocks across four two-week challenges, all with the same instructions.

The results were sobering:

• The combined portfolio lost about a third of its money.
• Out of 32 results, only 6 finished with a profit.
• The models traded very differently from each other on identical instructions. In one round, one model placed 1,418 trades while the best performer placed 158.

The founder of the contest summed it up: handing an AI model money and letting it trade on its own "is not a thing yet."

That doesn't mean AI is useless for trading. It means AI needs the right job. To see why, it helps to know how these AI models are built.

How today's AI is built (in plain words)

Most AI assistants today are built on a design called a transformer (introduced by Google researchers in 2017). Here is what that means for a trader:

• It learned from text, not from markets. The model was trained on a huge amount of writing — books, websites, articles, code — with one main goal: guess the next word. After billions of guesses it became very good at language, explanations and patterns in text.
• It writes one word at a time, by probability. When you ask a question, the model picks each next word from a list of likely words. That's why it sounds fluent — and also why it can sound sure while being wrong, and why the same question can get a different answer.
• It "pays attention" only to what's in front of it. The transformer's key trick, called attention, lets it weigh every word in your message against every other word. But it can only look at what fits in that window — not the whole market, and not anything that happened after its training unless you hand it in.
• Its knowledge is frozen. What it "knows" comes from training data that stops at a cutoff date. Live prices, today's news and your stock's last 5 minutes are invisible to it unless a tool feeds them in.
• It doesn't calculate the way a calculator does. It predicts what an answer should look like. It can often do math, but it can also produce a number that looks right and isn't.

In short: a transformer is a brilliant pattern-matcher for language. Trading needs live data, exact math, instant reactions and the same decision every time — the things this design was never built for. That's why pure AI trading is not reliable on its own, and why the limits below exist.

8 limits of AI in trading

1. AI doesn't see the market live

An AI model learns from data up to a cutoff date. After that, it only knows what it is handed in the conversation. If the price data it gets is a few minutes old, its answer is a few minutes old. In a fast market, that's a different market.

2. AI is slow for trading decisions

"Thinking" takes time — usually several seconds, sometimes much longer for complex questions. Markets move in milliseconds. By the time an AI finishes reasoning about a breakout, the entry you asked about may be long gone.

3. AI can only hold so much at once

An AI model has a limited "working memory" (often called its context window). It cannot take in every trade for thousands of stocks across many timeframes. When you give it too much, it summarizes, drops details, or misses things buried in the middle.

4. AI can't predict the future

No one and nothing can predict prices reliably — and AI is no exception. AI can sound very confident, but a confident answer is not a correct one.

5. AI isn't consistent

Ask an AI model the same question twice and you can get two different answers. Research has shown this happens even with settings meant to make answers repeatable. A trading rule must behave exactly the same way every time; an AI's judgment does not.

6. AI can make up numbers

AI can invent a price level, misread a chart, or get arithmetic wrong over large amounts of data — and present it smoothly. Every number it gives you needs a real source.

7. AI has human-like habits

AI models learned from human writing, so they can repeat human mistakes: overtrading, chasing a story, reacting to the last thing they saw. The contest above showed exactly that — some models traded far too often.

8. AI has no built-in sense of risk

AI doesn't feel a loss. It won't stop itself after three bad trades or refuse a position that's too big. Risk limits have to come from you, as hard rules.

Bottom line: relying on AI alone will not improve your trading odds. Used the right way, it can make you faster, more organized and more disciplined.

What AI is actually good at

• Turning your strategy into exact rules. You describe your setup in plain words; AI turns it into a precise rule and points out the parts that are vague ("what counts as a breakout — a touch, or a close above?").
• Explaining things. What a level is, why a gap matters, what a chart pattern means — in plain language, in any language.
• Reading fast. Summarizing news, company filings and earnings calls in seconds.
• Reviewing your trading. Going through your trade journal to find habits: when you break your own rules, which setups you trade best, what time of day you struggle.
• Stress-testing a plan. "What happens if it gaps over my entry?" "What if volume is light?"
• Brainstorming ideas to test — not ideas to trade blindly.

The better setup: AI writes the rule, the rule does the trading

The pattern that works is a simple split:

1. You decide what you want to trade and how much you're willing to risk.
2. AI helps you turn that idea into an exact, written rule.
3. A rule engine checks that rule against live data — instantly, and the same way every time.
4. You test it first: backtest on past data, then paper trade with fake money, then use small real size.
5. You stay in control of every decision and every risk limit.

AI does the thinking work it's good at. The rule does the fast, consistent work AI is bad at.

The data that helps AI make better calls

AI does its best work when it gets small, fresh, pre-calculated facts with their history — not raw price ticks. Pre-calculated facts fit in its limited memory, arrive instantly, and leave far less room for made-up numbers. These are the kinds of data that help most:

1. Odds from history — checked twice. "When price did this at this kind of level before, what happened next, and how often?" The best odds are measured over many years and only trusted when two separate time periods agree. Real outcomes replace guesses.
2. Key price levels, already marked. Prior day high, low and close; premarket and overnight highs and lows; the opening range; VWAP; pivots; round numbers; supply and demand zones.
3. How price behaved at each level before. How many times it was touched, held or broken; how many breakouts failed; whether old resistance turned into support.
4. Real buying vs. selling from every trade. Who was actually buying from sellers and selling to buyers — not a guess based on candle color.
5. Who is trading. Small orders vs. large orders vs. block trades, and how fast trades are coming in.
6. What's normal for the time of day. Typical volume, spread and number of trades at 9:35 vs. 1:00 PM, so "unusual" is measured, not felt.
7. Typical move size for each stock. How far its normal up and down moves go and how long they last — so targets and stops fit the stock.
8. Swing structure. Higher highs and lows, trend direction, channels, and how healthy the trend is, across several timeframes.
9. Range context. Average daily range, today's range so far, and how big a candle is compared with normal.
10. Candle shape tags. Inside bars, hammers, wide-range candles — labeled, so AI doesn't have to guess from pixels.
11. Session context. Earnings day, gaps, trading halts, news and its tone.
12. Your own results. Your trades and backtests, so AI can spot your habits.

How to tell AI about your setup

The quality of AI's help depends on how clearly you describe your setup. A good description includes:

• Timeframe — 5-minute chart? Daily?
• The exact trigger — "a 5-minute candle closes above yesterday's high," not "it breaks out."
• What proves you wrong — where the idea is invalid.
• Risk per trade — how much you'll lose if it fails.
• Which levels you use — prior day, opening range, VWAP, zones.
• Time window — only the first hour? All day?
• What to skip — earnings days, stocks under a certain price, low volume.

Weak request: "Is NVDA going to go up tomorrow?"

Strong request: "On the 5-minute chart, I look for a stock that opens below yesterday's high, then closes a 5-minute candle above it with volume higher than normal for that time of day. I'm wrong if price closes back below yesterday's high. I risk 1% per trade and skip earnings days. Turn this into an exact rule and list anything unclear."

The first asks AI to predict — something it can't do. The second asks AI to organize your thinking — something it does well.

What SEPT offers

SEPT Investing is a trading strategy automation platform. You describe your strategy in plain words; GOAX, SEPT's AI assistant, helps turn it into a rule; then you can scan, alert, backtest, paper trade and automate it on SEPT's own market data, built from every trade. That data is the kind of pre-calculated information described above:

1. Key levels with their history — a library of levels for each stock with touch counts, failed breakouts, support-turned-resistance flips, reaction strength, importance and nearby round numbers.
2. Real buying vs. selling from every trade — on every timeframe from 1 minute to a full year, split by small, medium, large and block trades.
3. Typical move size for each stock — how far its usual up and down moves go, how long they last and how often big moves happen.
4. What's normal for each time of day — usual volume, spread and trade count for each time slot and session.
5. Opening range profile — each stock's usual 5-, 15- and 30-minute opening range and opening volume.
6. Session levels — overnight, premarket, morning, midday, afternoon and after-hours highs and lows, prior day, pivots and VWAP.
7. Swing structure — swing highs and lows on 9 timeframes, with trend, channels and trend health.
8. Supply and demand zones and volume profile — where the most trading happened.
9. Candle shape tags and volume spike levels.
10. Stock facts — 52-week and all-time highs and lows, float, average volume and average range.
11. 400+ ready-made fields shared by the screener, alerts, backtesting, replay and GOAX.
12. Trading odds tables — in research now. SEPT research is counting what price actually did over ten years of 1-minute data every time a defined situation happened — for example, a stock touching yesterday's close and turning, or breaking its opening range. Each result is split into two periods (2017–2021 and 2022–2026) and marked validated only when both agree. Ideas that failed stay on record with the reason. These tables are not a live feature yet.

What SEPT's own research found

Hypothetical results on past data, 2017–2026. Past results do not promise future results.

• Nobody's pre-market signal tells you up or down. Every pre-open condition SEPT research tested landed between 49% and 51% for the day's direction — a coin flip. This is limit #4 above, measured.
• Big days are usually followed by smaller days. Across 1.51 million stock-days, after a day that moved 1.5 times the stock's average daily range or more, the next day was smaller 87.6% of the time (85.5% and 89.0% in the two periods).
• Some ideas didn't work, and we say so. SEPT research tested 29 trading ideas. Several — including breakouts at key levels and 19 event-based setups — showed no edge after costs. They stay on record, marked "tested, not profitable."

That is exactly the kind of checked, honest fact an AI assistant needs: something measured, not something guessed.

Use it with your own AI — coming soon. SEPT's API and MCP access, which will let AI assistants you already use read SEPT's pre-calculated data directly, are coming soon and planned to be ready before the end of 2026.

Data and tools help you see more clearly. They don't guarantee results, and past outcomes don't promise future ones.

See plans and start free → — Free tier available. Beta Pioneer: $29.99/month, available until Jan 2, 2027.

FAQ

Not financial advice. This page is for education only. Research figures are hypothetical results on past data (2017–2026) and do not include every real-world cost. Trading involves risk, including loss of money.

Sources

• Alpha Arena results (Nof1, 2026, US tech stocks): Business Standard / Bloomberg, “AI bots auditioning for Wall Street trading are mostly losing money”; ThinkAdvisor; Briefs.co summary (Nov 7, 2026); KuCoin News (trade counts).
• “Non-Determinism of ‘Deterministic’ LLM Settings” (arXiv 2408.04667)
• “Quantifying non-deterministic drift in large language models” (arXiv 2601.19934)
• SEPT research on U.S. stock market data, 2017–2026 (hypothetical results on past data).