What AI Can And Cannot Do In Trading In 2026

I write about AI in trading as someone who uses it daily and as someone who is exhausted by the discourse around it. The bull case from the AI-trading content economy in 2026 is roughly that we are months away from agentic models running pr

I write about AI in trading as someone who uses it daily and as someone who is exhausted by the discourse around it. The bull case from the AI-trading content economy in 2026 is roughly that we are months away from agentic models running profitable strategies autonomously. The bear case from the cynics is that the entire field is theatre and nothing has changed. Both are wrong in interesting ways. The honest answer is in the middle, it is narrow, and it is more useful to a working trader than either pole. AI in 2026 is a transformative workflow tool and a near-useless predictor, and the difference is the entire game.

A frame for the rest of the essay

It helps to start with a distinction the discourse refuses to make cleanly: there is a difference between using AI to do trading work and using AI to do trading. The first is now genuinely revolutionary in narrow ways. The second is still mostly nonsense. Most arguments about AI in markets collapse the moment you separate them.

When I say "AI" in this essay I mean the current generation of frontier large language models — Claude, GPT, Gemini and their peers — plus the smaller specialised models that have emerged around them. I do not mean classical machine learning, which has been deployed inside hedge funds for two decades and has nothing to do with the chatbot revolution. The two get conflated constantly in retail content and the conflation does no one any favours.

What AI is genuinely good at

There are five things the current generation of models does well enough that any trader ignoring them in 2026 is leaving real productivity on the table.

One: drafting and refining strategy code. Pine Script, Python backtesting frameworks, MQL, the ad-hoc helper scripts every discretionary trader needs. The model is a fast-typing junior with broad syntax recall and no intuition for whether the strategy is any good. That trade is fine if you keep it on the right side of the line. Writing Pine Script with Claude and GPT goes deep on the workflow.

Two: summarising research, earnings filings, and macro releases. A model that can ingest a fifty-page transcript and surface the three quotes that matter saves you an hour. It will hallucinate occasionally, so you verify. The verification is faster than the original reading would have been.

Three: structuring journal data. The model is excellent at taking an unstructured trade note and producing structured fields you can query later. This is not glamorous but it is the kind of compounding workflow improvement that pays dividends for years.

Four: backtest design and edge-case generation. Ask a model what conditions would break your strategy and it will produce a list that is roughly seventy percent useful. The thirty percent that is wrong does not matter because the seventy percent contains things you would not have thought of.

Five: risk-management calculators and position-size widgets. Ask the model to produce a Streamlit app that takes account size, R-percent, instrument, and entry-stop distance and produces a position size. You get something usable in twenty minutes. You used to spend half a day on this.

What AI is genuinely bad at

Now the other side, which the AI hype economy has every incentive to under-discuss.

One: live discretionary execution. The latency is wrong. The reasoning is too slow. The hallucination rate is non-zero. None of those are acceptable when the model is the thing pulling the trigger. There are narrow algorithmic contexts where models can sit in the loop with strong guardrails. Discretionary live trading is not one of them.

Two: novel pattern discovery on raw price action. Models are trained on text. They have no native intuition for OHLC bars or order-flow sequences. Multimodal models can see chart images and produce confident-sounding analysis, but the analysis is closer to autocomplete-of-trading-content than to genuine technical insight. It pattern-matches to what trading articles tend to say about charts that look this way, which is not the same as being right.

Three: sentiment in nuanced contexts. A model can tell you whether a Fed statement is more or less hawkish than the previous one. It cannot reliably tell you how the bond desk at Goldman is going to interpret the third paragraph of the dot plot relative to the consensus going in. The latter is what moves the market. The former is a summary.

Four: anything requiring real-time market awareness. Frontier models do not natively know what happened in the last sixty seconds. Tool-use can backfill some of that, but the model is then a router for tools, not a market participant. The illusion of real-time intelligence is exactly that.

Five: anything where a confident hallucination costs money in the next minute. This is the meta-failure mode. The model will produce a confident wrong answer with the same affect as a confident right answer. In a domain where the cost of wrong is asymmetric and immediate, that affect mismatch is dangerous. Code review can catch a hallucinated function name. Live trading cannot catch a hallucinated stop level.

Why AI live trading bots blow up goes into the failure-mode taxonomy in detail. The short version: most retail AI trading bots fail for boring, well-understood reasons that do not require a hot take to identify.

The middle ground that most retail content ignores

There is a category between "great use" and "bad use" that almost no one talks about, and it is where most of the interesting work for retail traders actually lives in 2026.

Strategy hypothesis generation. A model is bad at deciding whether a strategy works but useful at producing a wide list of hypotheses worth testing. Treat the output as a brainstorm input, not as an answer.

Backtest interpretation. A model is bad at running the backtest but useful at reading the output and asking "what would I be worried about here?" Treat the model as a sceptical reviewer. It will catch some things you would have missed and miss some things you should not have.

Bias auditing. A model that has read your last fifty journal entries can spot recurring patterns in your own behaviour faster than you can. The model has no skin in your ego. That is a feature for this specific task.

Documentation and playbook drafting. Your written strategy playbook should be specific, exhaustive, and boring. Models excel at producing specific, exhaustive, and boring text. Use it.

Education and mental model construction. Models can patiently explain a concept five different ways until one clicks. For learning, this is genuinely transformative. For decision-making, it is overrated.

LLMs as research assistants, not traders goes deep on the framing of "model as junior analyst, not principal." That framing is the single most important mental shift for using AI productively in trading work.

Why so much retail AI trading content is bad

Here is a frustration. The discourse around AI trading is dominated by two failure modes. The first is breathless hype from people selling AI signal services who have every commercial incentive to overstate what models can do. The second is performative cynicism from people who tried AI once, got a bad answer, and concluded the whole field is theatre. Both are wrong in ways that matter.

The hype side ignores that "this model can write code that backtests positively on historical data" is not the same claim as "this model can generate live alpha." The cynic side ignores that the workflow productivity gains are real, large, and compounding even if the alpha-generation claim is wrong.

What is missing in both camps is the boring middle: a working trader using AI as a sharp, fast, occasionally wrong assistant on a specific set of tasks, with discipline about which tasks. That is what the next decade of serious retail trading actually looks like, and it is much less interesting as content than either pole.

The model is brilliant at writing the code that runs your strategy and useless at telling you whether the strategy is any good. Conflating those two capabilities is the most expensive mistake in retail AI trading right now.

— Working notes, AI-trading research review, Tradoki, Q4 2025

How to think about this for the next twenty-four months

Two predictions and a recommendation.

Prediction one: the next two years will see meaningful improvements in agentic frameworks for trading research workflows. Models orchestrating tools, running backtests, summarising results, iterating on strategy code. This will compress what currently takes a week of part-time research into a day. That is large. That is real. It is also still research workflow, not edge.

Prediction two: the gap between "AI can produce a strategy that backtests well" and "AI can produce a strategy that is profitable live with size" will not close meaningfully in twenty-four months. The gap is not a model problem. It is a market microstructure, regime-change, and execution-cost problem, none of which are bottlenecked on model capability.

Recommendation: build your AI-in-trading toolchain assuming the workflow gains are permanent and the prediction gains are not coming. Optimise for fewer hours spent on grunt work, faster iteration on hypotheses, better-organised journal data, more thoroughly documented strategies. Do not optimise for autonomous decision-making. The first set compounds in your favour. The second set is a trap.

The disclosure I want to make explicitly

I am writing this from a platform that uses AI in our own workflow extensively. Our cohort materials are partly drafted with AI. Our research summaries are accelerated by AI. Our internal tooling is built faster because of AI. None of this means we believe AI is going to predict the next move in EURUSD. We use the technology where it works and we are aggressive about excluding it from the places it does not. That selectivity is the whole point of this essay.

If you take one thing away, take this: in 2026, the trader who is going to win is the one who has integrated AI deeply into research and process while remaining absolutely disciplined about not letting it touch the decision moment. The trader who lets AI make the decision will lose to the market. The trader who refuses to use AI at all will lose to the trader who uses it well. Those are the two failure modes. The third option — selective, disciplined, integrated — is the one worth building toward.

● FAQ

Can AI predict the market?

No. Not in the sense the question is usually asked. Models can summarise, structure, code, and surface patterns in defined datasets. None of that is prediction in the way retail content makes it sound. Anyone selling AI market prediction is selling something a serious quant would not buy.

What is AI genuinely good at for traders right now?

Five things: drafting and refining strategy code, summarising research and earnings filings, structuring journal data, accelerating backtest design, and building risk-management calculators. Notice these are workflow accelerators, not edge sources.

What is AI genuinely bad at?

Live discretionary execution, novel pattern discovery on raw price action, sentiment in nuanced contexts, anything requiring up-to-the-second market awareness, and anything where confident hallucinations would cost money in the next sixty seconds.

Are AI trading bots viable?

Most retail AI trading bots blow up. The reasons are well-documented: lookahead bias in training, regime change, brittle prompts, and the gap between historical and live execution. We have written about this in detail elsewhere.

Should retail traders use AI at all?

Yes — but as a research and process tool, not as a decision-making oracle. The trader who treats Claude as a fast-typing junior analyst captures most of the upside. The trader who treats it as an edge source captures most of the downside.