AI Trading: What Works, What Does Not, and How to Start

AI can help with research, analysis and even sending orders, but none of that proves a profit. That is the honest starting point for AI trading in 2026. The hard job is to tell a useful tool from a marketing label, and to check what any tool gives you.

This guide is written for a trader who wants to use AI without being sold on it. You will see what AI is good at, where the evidence is weak, how to check the output of any tool, and which questions to ask before you pay for an “AI” product.

AI trading in 30 seconds

• What AI trading is: using a language model or a machine-learning tool to research markets, analyze charts and companies, help you code and test a strategy, or send orders.
• What the evidence shows: good results on some narrow tasks, such as reading financial statements. Using AI does not by itself show that a strategy is profitable after costs.
• Where AI helps most today: reading filings, transcripts and news; writing Pine Script or Python; describing a chart.
• What to watch: fees, slippage and live execution, which no backtest or demo can guarantee.
• The AI-label question: ask what a paid “AI tool” adds beyond a public chat subscription and your existing chart software.
• First check for any AI output: open the source it cites, match the numbers word for word, then decide.
• First place to start: a general chat assistant such as ChatGPT, Claude or Perplexity, together with the tools you already use (TradingView, your broker).

What is AI trading?

AI trading means using machine-learning or language models for a trading task you would otherwise do by hand. The three common tasks are research, signal and strategy work, and execution. They are very different jobs and need very different checks.

1. AI drafts, you checkResearchAI summarizes a filing, transcript or news. Check: open the source and match every number.
2. AI drafts, you checkSignals and strategyAI marks chart levels or writes a rule. Check: test the rule on data it has not seen.
3. Software sends ordersExecutionA program sends orders when a signal fires. Check: paper-trade first, cap size in dollars.

The dark line marks the hand-over: from the third step on, software acts on your account. Graphic by Kagels Trading.

In practice most retail traders use AI for the first job and sometimes the second. Research means letting a model read a 10-K, an earnings call transcript or a news stream and summarize what matters. Signals and strategy work means asking the model to draw levels on a chart, write a short screener, or build and test a rule. Execution means sending orders by code. AI may produce the signal, but whether a program should press the button is a separate decision.

Research, signals and execution

Each step in the graphic above has its own check. For research, open the source document and match every number. For signals and strategy, a rule that worked on past prices must also hold on out-of-sample data. For execution, the risk moves from the model to the plumbing: before software sends live orders, paper-trade it, cap size in dollars, and review the first live week trade by trade.

Execution is a separate decision, and this guide treats it that way. The first two steps can be checked at your desk. The third one acts on your account, so the questions in the section on AI trading bots below come first.

AI versus rule-based automation

AI is not the same as algorithmic trading, but the two overlap. Algorithmic trading uses a program to decide whether, when and how to trade. Many algos run on human-written rules, but algorithmic trading can also use machine-learning models that learned patterns from data. The useful question is not “AI or algo” but “what exactly does this tool do, and can I check it?”

A machine-learning model is trained on data and finds patterns it was not explicitly told to look for. A language model can read a filing. A tree model can learn that two indicators behave differently in a trending market than in a range. Automated trading describes how the order is delivered; the signal behind it may come from a rule, from a model, or from both.

Three practical points follow from this difference. First, “AI trading” in a product name does not tell you what is inside. Second, using AI to write a rule-based strategy does not turn the strategy into an AI model. Third, many tools on sale today mix rules and machine-learning parts.

AI claims: what the label tells you

The word “AI” on a product page does not tell you what the product does. The U.S. Securities and Exchange Commission charged two advisers in March 2024 with exactly this problem. Delphia (USA) Inc. had told clients it put “collective data to work to make our artificial intelligence smarter,” and Global Predictions Inc. had called itself the “first regulated AI financial advisor.” The SEC found that neither firm had the AI capabilities it advertised and that both had violated the Marketing Rule. The two settled for a combined $400,000 in civil penalties (Delphia $225,000, Global Predictions $175,000) without admitting or denying the findings (SEC press release 2024-36).

Some AI products are wrappers, and that is not automatically bad. A wrapper is a product built on top of a public language model from another company, with its own interface, prompts or data around it. A good wrapper can save you time, bundle data or connect to your broker. What matters is what you get for the extra fee.

Before you pay for an AI product, ask four plain questions. They work for a chat add-on, a scanner or a bot:

• Which model does the product use, its own or one from another company?
• Does the vendor show dated, reproducible test results, not just a chart?
• Can the vendor explain what data the tool uses and how results were measured?
• What does it add beyond a public chat subscription and your existing chart software, for example a dataset, a broker connection or a workflow you would need hours to build?

The answers tell you what you are paying for, not whether the tool works. A clear answer on the model and the data is a start. Whether the tool helps you still has to be shown on your own tasks, with your own checks.

Does AI trading work?

It depends on which job you are asking about. Four things get mixed together in that question: time saved on preparation, prediction accuracy on a defined task, simulated results of a strategy, and net returns in a funded account. The evidence for each is different, and one does not prove the next.

What AI is good at: research and analysis

On one defined analysis task, a language model beat the analyst consensus. A University of Chicago Booth working paper by Kim, Muhn and Nikolaev gave GPT-4 anonymized financial statements and asked whether earnings would go up or down. With a step-by-step chain-of-thought prompt, GPT-4 reached an accuracy of about 60.4%. Analysts’ forecasts made one month after the earnings release reached a median of about 52.7% (authors’ policy brief on the working paper, SUERF, October 2024). This is a narrow prediction task on anonymized filings. It does not show that trading on these predictions is profitable.

For daily work, the practical gain to look for is time. On a long quarterly report, a careful prompt plus a human cross-check against the filing can be faster than reading everything start to finish. How much faster depends on your task and your checks, so measure it on your own work before you pay for a tool.

Where the evidence is weak: autonomous trading

A public benchmark of autonomous AI traders did not show reliable trading skill. A benchmark by Fan et al. (AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial Markets, arXiv 2512.10971, version 1, 1 December 2025) let six language models trade US stocks, Chinese A-shares and crypto in October and November 2025. The agents searched for real-time information on their own. The paper calls the setup live, but its abstract says the trading frequencies serve “to simulate live financial environments”, and the paper does not name a broker, a starting capital, or how fills and costs were calculated. Its finding: “general intelligence does not automatically translate to effective trading capability”, with most agents showing poor returns and weak risk management.

Read this as a warning about autonomy, not as a measure of real-money losses. The test ran for weeks, not years, and the paper does not document funded execution, so its returns are not proven account results. It says nothing about your own workflow, where you check the output before anything is traded.

A separate line of research asks what happens when many AI traders learn at the same time. In “(Deep) Learning to Trade: Co-learning Bias and Market Outcomes” by Sangiorgi, Gufler and Tarantino (SSRN 5375160; authors’ version, August 2026), AI traders learn by deep reinforcement learning in a simulated market. Each one mistakes part of the others’ price impact for its own and trades less aggressively than the standard game-theory benchmark (the Nash equilibrium). In this simulation, the result is higher collective returns for the AI traders than in that benchmark, but lower market efficiency, meaning prices carry less information.

A real-market study looks at retail traders and ChatGPT. Xiaoxue He of Lancaster University Management School (“When ChatGPT Stops Talking”, SSRN 6350140, version September 2026) uses 14 unexpected ChatGPT outages as a natural experiment. When ChatGPT is available, retail order flow is more correlated across stocks, and this common part “shows little evidence of return predictability.” The effect on how informative trades are runs in opposite directions: retail buying becomes more informative, retail selling less informative. Liquidity also moves more in sync across stocks. The two papers are separate works with different methods, so neither confirms the other.

Why backtest results often do not survive

A good backtest on old data does not promise a good result on new data. The most-discussed cause is overfitting: a strategy is tuned to past prices so closely that it no longer describes a repeatable pattern. A backtest can also miss because the market has moved into a different regime, because costs have changed, because the data had survivorship bias, or because live fills differ from the fills the test assumed.

Two checks catch many of these problems. First, split the data: build the rule on one part and test it on a part the model has not seen. Second, run a walk-forward test that moves this split through time. Backtests can model spread, fees and slippage; they cannot promise that a live account will match those assumptions. Our backtesting guide walks through the full procedure.

Some research results are positive, but they are narrow. A paper on arXiv in October 2025 by Anic, Barbon, Seiz and Zarattini (ChatGPT in Systematic Investing) reports that an LLM-filtered momentum strategy on the S&P 500 improved its out-of-sample Sharpe ratio from 0.79 to 1.06 after transaction costs. The result applies to a long-only equity momentum strategy, assumes costs of about 2 basis points per trade, covers a short window after the model’s training cutoff, and the authors note that the raw alpha only clears a 10% significance level. It is a research result on a tightly defined test, not a product.

AI trading tools: pick by task

There is no single best AI trading tool, only a better tool for each job. Decide which job you want to improve this month and start there. Research is covered by the general chat assistants. Chart analysis is covered by TradingView’s AI add-on if you already pay for TradingView. Scanners sit with specialist vendors. Execution is a separate category, judged with the questions in the section on bots below.

Job	Example tool	Check this output
Read a document fast	ChatGPT, Claude, Perplexity, NotebookLM	Numbers in the summary match the source word for word
Explain a chart	TradingView AI Copilot	Levels and signals match price history and your own reading
Scan or screen the market	ChatGPT, Claude, Trade Ideas	The scan returns the stocks that actually match your rules
Draft and test a rule	Claude Code, ChatGPT Code Interpreter	The result holds on data the rule has not seen
Trade on a signal	A broker-connected program	Live trades match the strategy on paper and on fees

Only the last job sends orders to your account. The first four stay at your desk, even when the same vendor also sells order automation, as Trade Ideas does in its Premium plan.

Prices and features change often, so check them on the vendor’s own page. The notes below were checked on the dates given. Plans, prices and availability can differ by country.

• General chat assistants. ChatGPT, Claude and Perplexity each have free and paid plans. Compare current plans on the ChatGPT, Claude and Perplexity pages. NotebookLM from Google answers from the documents you upload, and its answers still need checking against those documents.
• TradingView AI Copilot. It is an add-on to a paid TradingView plan. It works with Essential, Plus, Premium or Ultimate, not with the free Basic plan. Karsten bought it from Germany at €7.50 before VAT in September 2026; the US checkout price was not checked, so look at the price in your own account or on the TradingView pricing page.
• TrendSpider. Its pricing page lists a machine-learning model lab and an AI assistant called Sidekick. The lab names KNN among its models. Standard plans include a small monthly allowance of Sidekick messages; more messages cost extra. Promotions change often, so check the TrendSpider pricing page.
• Trade Ideas. Billing monthly or yearly gives different prices. On 25 October 2026 the Trade Ideas pricing page listed Standard at $127 billed monthly, or $1,068 billed annually ($89 a month equivalent), and Premium at $254 billed monthly, or $2,136 billed annually ($178 a month equivalent). Premium lists AI signals and an auto trading feature with connected brokers; Standard lists in-app trading but not auto trading.
• Tickeron. It advertises AI models together with performance figures. Check every number against independent, dated records before you pay. If the vendor cannot show them, treat the claim as unverified.

What do you want to improve this month?

• Read filings or news faster→A general chat assistant (ChatGPT, Claude, Perplexity or NotebookLM)
• Explain and mark up charts→TradingView AI Copilot
• Scan the market by your own rules→Trade Ideas or a chat assistant
• Draft and test a strategy in code→Claude Code or ChatGPT Code Interpreter
• Not sure yet→A general chat assistant, used as a research partner before anything is bought

Execution is deliberately not here. Read the next section before you let any tool send orders.

AI trading bots: where execution changes the risks

An AI “signal” and an AI “order” are different things. A signal tells you to do something. An order is a specific instruction sent to a broker. A fill is what you actually got, after spread, slippage and fees. A backtest can model an alert as a fill at an assumed price and cost; in a live account the three can differ.

If you consider an AI trading bot, ask four questions first. They matter more than any accuracy claim or demo:

• Where does it run? On your machine, on your broker’s platform, or on the vendor’s server?
• What permission does your broker account give it? Market data only, read-only, or order placement?
• How is risk capped? Per trade and per day, as a dollar amount, not as a percentage of an untested forecast.
• What is the audit trail? Can you rebuild a losing trade step by step, with time stamps from the bot and from the broker?

A regulator has warned directly against AI trading-bot promises. For US-regulated derivatives, the Commodity Futures Trading Commission published a customer advisory in January 2024 titled “AI Won’t Turn Trading Bots into Money Machines.” It says AI technology cannot predict the future or sudden market changes, and it tells readers to consider what fees, spreads and subscription costs do to returns.

AI in practice: one task and how to check it

The worked example on this site is a published, checked test. It is Karsten’s TradingView AI Copilot review, published on 2 October 2026 after a test on 30 September 2026. He paid for the add-on, gave the Copilot seven tasks (two chart analyses, a Pine Script indicator, a bug fix, a screener query and two alerts), and documented every answer.

The review found four problems worth knowing. Each one was caught by a human check, not by the Copilot:

• On a DAX hourly chart, the Copilot drew seven levels and three trade plans. One trade plan had a calculation error on an ATR-based buffer, and a second had a reward-to-risk ratio that only fits an entry the plan did not name.
• On a daily gold chart, the Copilot picked a swing low as the lowest level but omitted a lower low at $3,943 on 30 June 2026, about $17 below the drawn line. That matters for stop placement.
• The Copilot wrote that about 70% of an October rate hike was priced in. On the same day, Yahoo Finance reported about 35% per CME FedWatch; the Copilot’s number was probably outdated (a rate probability, not the Fed funds rate itself).
• On a Pine Script Donchian Channel, the generated code compiled, but the breakout signal did not fire once in more than a year of gold. The Copilot diagnosed and fixed the bug only after the human noticed the missing signals.

The same method works for any first AI task. Pick one job you already do by hand, give the model the same inputs you would use, compare its output with your own, check every number against its source, and keep notes for the first week. Then decide whether the tool is worth its monthly fee.

To go further, turn an AI-written rule into a proper test. The Pine Script guide covers the language the Copilot writes, and the backtesting section above shows the two checks that come next.

Risks and misleading offers

AI did not invent trading fraud, but it changed what fraud can look like. Four patterns are worth knowing before you move money, whatever the marketing says.

Celebrity deepfakes are cheap to make. A short video of a well-known person recommending a trading platform can be generated with little effort. Check the person’s own verified channel; if the video does not appear there, treat it as unverified.

Chatbots can look like a sales team. A contact who answers instantly, around the clock, in several languages and with very uniform wording may be automated messaging rather than a licensed representative. Confirm the firm’s registration in a public register before you send money.

Fraudulent AI “platforms” are a known pattern. The CFTC advisory describes Mirror Trading International, which took bitcoin investments starting at $100 and promised returns of at least 10 percent a month from a trading bot. According to the CFTC, very little money was actually traded; the operation was a Ponzi scheme.

AI stock-signal apps that bill monthly need the same checks. The European Securities and Markets Authority (ESMA warning, March 2025) writes: “There is a growing number of websites and apps that offer AI-generated trading ideas and suggestions, often in exchange for expensive monthly or yearly fees, using names such as ‘stock picking’ or ‘stock signal’.” It adds that “no one and nothing, including AI tools, can accurately predict financial markets.” A monthly fee alone does not make an app fraudulent, but unverifiable performance and guaranteed returns are warning signs.

Conclusion: a sensible first task and its checks

Treat AI as a research assistant whose work you check, not as an autonomous trader. Any benefit should show up in your own notes: time saved, fewer errors, more ideas tested. If it does not, the tool is not worth its fee for you.

A sensible first task takes about an hour this week. Pick the next earnings release on a stock you already follow. Download the transcript. Ask a chat assistant to summarize the three numbers that moved and the three risks management flagged. Open the transcript and the previous quarter’s release and check each sentence. If the model got everything right, use the same workflow next quarter. If it invented a number, you have learned how far to trust it on this stock.

Spend the second week on charts. Pick a chart you look at daily. Ask a chat assistant, or the TradingView AI Copilot, to describe the current trend, the two most important support and resistance levels, and the one indicator it would use next. Compare its output with your own. If it agrees too easily, keep checking.

Letting AI send orders is a separate decision. Postpone it until the first two steps have run long enough that you know what you would automate. Better tools exist than ever before, but there is still no autopilot and no guaranteed return.

This guide is education and market analysis, not investment advice. It does not recommend any product, strategy or trade for your own situation, and any tool named here can cost you money if you use its output without checks.

Frequently asked questions about AI trading

Can you make money with AI trading?

Using AI does not by itself show that you will make money. AI can help with research, chart analysis and even order execution. Whether a strategy is profitable after fees and slippage has to be shown with your own dated records, and no AI tool can promise it.

Does AI trading guarantee profits?

No, and promises of guaranteed returns are a warning sign. The CFTC advisory says AI technology cannot predict the future or sudden market changes, and ESMA writes that “no one and nothing, including AI tools, can accurately predict financial markets.” Treat any product that promises guaranteed returns with suspicion.

Is AI trading the same as algo trading?

The two overlap but are not the same. Algorithmic trading means a program decides when to trade, using rules a human wrote, models that learned from data, or both. AI trading usually refers to the models that learned from data, or to language models used along the way. Many systems mix the two, so ask what the specific tool does.

Do I need to code to use AI for trading?

You do not need to code for research or chart analysis. A general chat assistant is enough for that. For a simple screener or backtest, code assistants such as Claude Code or ChatGPT Code Interpreter can write and run small Python scripts from a plain-English description. The harder part is usually knowing which rule to test.

What are the best free AI tools for trading?

Free versions of the general chat assistants cover simple summaries. NotebookLM answers from documents you upload, and its answers still need checking. TradingView has a free Basic plan with charts but without the AI Copilot. Beyond that, “free” often means a free trial; check the renewal date and how to cancel before you give a card number.

Can an AI place trades for me?

Yes, through a broker connection, and some vendors sell exactly that. Live execution brings spread, slippage, fees and account rules that a backtest can model but not guarantee. If you try it, start with a paper account, which uses simulated fills, cap risk in dollars, and review the first live week trade by trade.