AI and trading
The Best Way to Use AI in Trading (and What It Really Costs)
By Alex J. Wang, Market Research, SEPT Investing · Published · Last updated
Before you plug an AI model into every candle, do the math.
A new kind of post is going around: "AI reads every candle and calls LONG, SHORT or WAIT — and it feels illegal." Usually there's a slick video, a paper account that went up a few percent, and a free script to copy.
The idea behind it is not crazy. Feeding AI clean, structured market data is a good idea. But letting the AI make a fresh decision on every single candle has three problems the videos don't show: the cost, the consistency, and the proof.
Problem 1: The bill
AI models charge by the "token" — roughly a short word. You pay for what you send in (the instructions plus the market data) and more for what comes back (the decision and the reasons).
As of November 2026, a top AI model lists at $8 per million tokens in and $40 per million tokens out. A typical "read this candle and decide" request is about 5,000 tokens in and 500 out. That works out to about 3 to 6 cents per candle, depending on how much of the prompt can be cached.
That sounds cheap. Now multiply it:
| Setup | AI calls per day | Cost per day | Cost per month (21 trading days) |
|---|---|---|---|
| 3 symbols, 4-hour candles (what most demos show) | 18 | about $1 | about $20–30 |
| 100 stocks, 5-minute candles | 7,800 | $230–$470 | $5,000–$10,000 |
| 100 stocks, 1-minute candles | 39,000 | $1,100–$2,300 | $24,000–$49,000 |
Estimates using list prices as of Nov 2026, regular trading hours only (390 minutes a day). Models set to "think harder" write more and cost more.
Demos look cheap because they watch three symbols on slow candles. Turn the same idea into a real day-trading scanner and the AI bill can be bigger than the account.
Problem 2: Same chart, different answer
AI models write their answers word by word with some built-in randomness. Ask the same question twice and you can get two different answers — research shows this happens even with settings meant to make answers repeatable.
That's fine for brainstorming. It's a serious problem for trading. A trading rule has to do the same thing every time it sees the same situation, or you can never test it, trust it, or improve it. If the AI says LONG at 10:05 and WAIT on the identical setup at 10:20, what exactly are you testing?
And every decision takes seconds. A rule engine reacts in a fraction of a second.
Problem 3: A few good trades isn't proof
Most of these demos show a paper account over one or two weeks — often fewer than ten trades, sometimes on a replay the author picked. Seven winning trades in a row can happen by luck alone.
What real testing looks like:
- Many years of data, not two weeks.
- Two separate periods that agree. SEPT research splits every test into 2017–2021 and 2022–2026 and only trusts a result when both periods show the same thing.
- Costs included — commissions, slippage and, now, AI tokens.
- Failures kept on record. SEPT research tested 29 trading ideas; several showed no edge after costs, and they stay on record marked "tested, not profitable."
When you see "AI made 9% in two weeks," ask: how many trades, over how long, after costs, and does it hold up on a different period? If the post can't answer, it isn't proof — it's a demo.
Where AI is strongest: fundamental analysis
AI does its best work where speed doesn't matter and the input is text. That's fundamental analysis. Earnings reports, 10-K and 10-Q filings, earnings call transcripts, guidance changes, insider filings, analyst notes, industry news — this information changes in days or weeks, not milliseconds. Nothing is lost if the AI takes 20 seconds to think.
Good fundamental jobs for AI:
- Summarize an earnings report in plain words: what beat, what missed, what changed in guidance.
- Read a long filing and pull out risks, debt changes, share dilution or new lawsuits.
- Compare companies in the same industry on growth, margins and valuation.
- Track a story over time: "How has management's tone on demand changed over the last four calls?"
- Explain why a stock moved using the news and filings from that day.
This is the work that used to take an analyst hours. AI can do a solid first pass in minutes, at a cost of cents — and it doesn't need live market data to do it.
Technical analysis: AI can help, but not on raw live data
AI understands technical analysis well. It can explain support and resistance, chart patterns, volume, trends and setups, and it can help you design a technical strategy. Where it breaks down is reading raw, real-time market data and making the call itself. Two reasons:
- Cost. Live prices never stop. Sending every candle of every stock to an AI model is the expensive part — the bill in the table above.
- The AI's own limits. It can only hold so much data at once, it takes seconds to answer, it can give different answers to the same chart, and it can get arithmetic wrong over thousands of numbers. Raw candles are a lot of numbers.
The fix isn't to give up on AI for technical analysis. It's to stop feeding it raw data. Let fast, exact software do the math on every trade and candle — levels, volume compared with normal, buying vs. selling, ranges, trend structure — and give AI the finished results. Then AI does what it's good at: reasoning, explaining, and helping you write and improve the rules.
So what is AI actually good for in trading?
A lot — when it has the right job. Here are four ways to use AI that are useful, cheap and testable.
1. AI writes the rule (once)
You describe your setup in plain words. AI turns it into an exact rule and points out what's vague: "what counts as a breakout — a touch, or a 5-minute close above?" Then a rule engine runs that rule on every candle, instantly and the same way every time. Cost: one conversation, well under a dollar.
2. AI explains, only when something happens
Don't ask AI about every candle. Let the rule engine watch the market, and only when your rule fires, send that moment to AI: "explain what happened here and what would prove this idea wrong." Cost: if your rules fire 30 times a day, about $1–2 a day.
3. AI reviews your trading after the close
Once a day, give AI your trades and results. It's very good at spotting habits: you break your own rule after two losses, you trade worse after lunch, one setup carries all your profit. Cost: about one call a day — a few dollars a month.
4. AI helps you test ideas, not trade them
AI can turn "I think gaps on earnings day behave differently" into a precise test you can run on years of data. The data gives the answer, not the AI's opinion.
| Way to use AI | AI calls per day | Rough monthly cost | Same result every time? |
|---|---|---|---|
| AI decides every 1-min candle, 100 stocks | 39,000 | $24,000–$49,000 | No |
| AI writes the rule, rule engine runs it | ~0 (once) | under $1 | Yes |
| AI explains only when a rule fires | ~30 | $20–40 | Rule: yes |
| AI reviews your day after the close | 1 | a few dollars | — |
Start with computed data
For technical work, the best starting point is computed data. Whatever job you give AI, it does far better with small, ready-made facts than with raw prices. Pre-computed data fits in the AI's limited memory, arrives instantly, and leaves much less room for made-up numbers.
Instead of handing AI 390 one-minute candles and asking "what do you see?", hand it facts like:
- Price is 0.3% below yesterday's high, a level touched 4 times this month with 2 failed breakouts.
- Real buying vs. selling from every trade over the last 15 minutes: buyers ahead, mostly large orders.
- Volume is 2.1 times normal for this time of day.
- This stock's usual daily range is 3.2%; today it has already moved 2.8%.
- Earnings are in two days.
Now the AI isn't guessing from a picture. It's reasoning over measured facts — which is the job it's actually good at.
What SEPT offers
SEPT Investing is a trading strategy automation platform built around this split. You describe your strategy in plain words; GOAX, SEPT's AI assistant, helps turn it into an exact rule; then SEPT's rule engine scans, alerts, backtests, paper trades and automates it on SEPT's own data, built from every trade — no AI call per candle.
That data is the pre-computed kind described above: key levels with their history, real buying vs. selling from every trade, what's normal for each time of day, typical move size for each stock, swing structure on many timeframes, session levels, and 400+ ready-made fields. On the fundamental side, SEPT also collects SEC filings with plain-English summaries, earnings, insider and analyst actions, and market news. SEPT's API and MCP access — so the AI assistant you already use can read the same data directly — are coming soon, planned before the end of 2026.
Want the background on why AI struggles to trade on its own? Read AI Trading in 2026: What AI Can and Can't Do for Traders.
See plans and start free → — Free tier available. Beta Pioneer: $29.99/month, available until Jan 2, 2027.
FAQ
How much does it cost to let AI decide every trade?
At November 2026 list prices for a top model, about 3–6 cents per candle. That's around $1 a day for a few symbols on 4-hour candles, but $24,000–$49,000 a month for 100 stocks on 1-minute candles.
Is it a good idea to let AI call LONG or SHORT on every candle?
It's expensive at scale, slow, and not consistent — the same input can get a different answer. It works better to have AI write an exact rule once and let a rule engine apply it.
What is AI best at in investing and trading?
Fundamental analysis — reading earnings reports, filings, call transcripts and news, comparing companies and explaining why a stock moved. That work doesn't need real-time data, so AI's speed and cost limits barely matter.
Can AI do technical analysis?
Yes, it understands technical analysis well and can help you design a strategy. It struggles when it has to read raw, real-time price data and make the call itself — that's expensive and runs into its memory, speed and consistency limits. Give it computed data instead: levels, volume vs. normal, buying vs. selling, ranges and trend structure.
What's the cheapest useful way to use AI for trading?
Use AI to write and check your rules, explain signals only when they fire, and review your trades after the close. All three cost a few dollars a month or less.
How do I know if an AI trading result is real?
Ask how many trades, over how many years, after costs, and whether it holds up on a separate time period. A few weeks of paper trading isn't enough.
What data should I give AI for trading?
Small, pre-computed facts with history: key levels and how price reacted there, real buying vs. selling, volume compared with normal for the time of day, typical move size, and upcoming events like earnings.
Not financial advice. This page is for education only. Cost figures are estimates using published list prices as of November 2026 and will change. Research figures are hypothetical results on past data and do not include every real-world cost. Trading involves risk, including loss of money.