What does an AI crypto trading bot actually do?

Published: 12 November 2026 | Updated: 12 November 2026

On 7 November 2026, three crypto-automation tools describe three different things under "AI" in their own documentation.

• Cryptohopper's help centre says its "A.I. stands for: Algorithmic Intelligence". It describes scoring the strategies you supply by how often their past signals hit a profit target.
• 3Commas says it "does not provide trading advice or AI signals" and instead connects external AI tools to its bots.
• The open-source FreqAI module in Freqtrade trains models such as LightGBM and XGBoost and retrains them on a schedule.

None of these descriptions, on its own, tells you whether the bot makes money.

What can 'AI' mean in a crypto trading bot?

These mechanisms often coexist in one product. Selection and optimisation are forms of statistical learning in a broad sense. Calling something a scorer rather than a prediction model does not settle whether it belongs under the AI label, and neither description establishes profitability.

The useful question is what is learned from what data, and how is it checked on data it did not see?

We would also separate the assistant from the trading process. An LLM can help produce a strategy without being the model that generates its market signals. Likewise, a strategy can use parameters selected by backtesting while executing explicit rules.

What do the platforms' own documents say?

The table records documented mechanisms and the questions those descriptions leave open. It is not a ranking. Documentation can explain what a feature is intended to do without exposing its server-side implementation or establishing how it performs.

All vendor quotes above were re-checked against the live (or archived) pages on 2026-11-12.

Missing disclosure remains an open question. It is not evidence that a vendor has no machine learning, and it is not a reason to infer an undisclosed model.

Why can't a good backtest prove an AI bot works?

Our simplified experiments used public Binance spot BTCUSDT and ETHUSDT hourly data from 2022-01-01 to 2026-10-31. They demonstrate evaluation problems in our own models and simulators, not any vendor's system. A deliberate look-ahead bug made a broken test print 84.6% accuracy; the walk-forward model's AUC was 0.51, with intervals spanning chance. The latter test did not isolate retraining.

The prediction task used hourly candles resampled into 4-hour bars and a binary target: “up” meant the next 24-hour log return, over the next 6 bars, was above zero; zero counted as “down.” Inputs were historical returns, volatility, RSI, an exponential-moving-average gap and a volume z-score, computed from information available at the bar's close; a predicted probability above 0.5 meant “up.”

How a look-ahead bug made a test print 85% accuracy

A deliberately added centred 13-bar moving average used 6 future bars, leaking future information into logistic regression despite a chronological train/test split.

The resulting 84.6% accuracy and AUC 0.93 are what a broken test prints, not evidence of model skill; without that feature, the same model's accuracy was about 50%.

Why shuffled train/test splits inflate results

Neighbouring bars share most of their 24-hour label window, so observations are dependent whatever the split; separately, shuffling puts immediate neighbours into training, leaking overlapping test outcomes through their labels, and also permits training on later data.

The blocked folds train on both sides of each test block, with a 24h purge before and 48h embargo after, so only the walk-forward test trains strictly on earlier data.

This same-sample comparison measures the gap between split designs without apportioning it between overlapping-label leakage and training on later observations.

Was a walk-forward model distinguishable from chance?

Our logistic regression used a rolling 365-day window, refitting every 30 days to produce 35 models; its AUC intervals included chance at 0.5, and accuracy was below an “always up” guess.

This evaluated one walk-forward procedure and did not isolate the effect of retraining: that would require a frozen-model comparison or another refit schedule on the same dates, which we do not report.

A log-return is the natural log of ending value divided by starting value. 0 means no change, about +0.69 means the value doubled, and about −0.69 means it halved. So +0.59 is not +59% and −0.91 is not −91%. The model forecasts direction over the next 24 hours, but this illustration re-decides its position every 4-hour bar, so it tests one trading policy built on the forecast, not the forecast horizon itself.

How much did the best of 16 grid settings overstate the next week?

Our own spot-grid simulator selected the highest-return setting from 16 combinations of range width and grid count on the previous 7 days, then measured its next-7-day return; the range also came from the selection week, making the in-sample figure optimistic by construction.

The average gap below persisted under the alternative intrabar-path assumption, but it does not mean every week had a shortfall. The intervals come from a moving-block bootstrap that resamples runs of 12 consecutive weeks, so dependence between neighbouring weeks is kept.

Anny does not support grid bots.

Does the top-scoring strategy stay on top?

Our selector chose weekly among five hourly rules using completed trailing-90-day hit rates, where a hit meant reaching at least 2% above the signal close within 24 hours, then compared the chosen rule's next-28-day hit rate with the five-rule average.

There was a small observed difference—BTC slightly below and ETH slightly above that average—but no confidence interval for the paired difference was computed, and the evaluation windows overlap.

This does not show equivalence or that selection has no value, and hit rates are not profits: this setup had no stop, exit or costs.

These experiments cover simple features and models, a limited asset set and a single venue and period. Overlapping labels reduce independence, execution is simplified, and hit rates do not measure profit. These tests show how evaluation choices change results. They do not show that machine learning can't work in crypto, and they say nothing about any specific vendor's models.

What evidence should you ask an AI bot vendor for?

Use the vendor-evidence checklist below. A model description alone leaves performance unresolved; a return chart alone leaves the mechanism and testing procedure unresolved.

Treat an unanswered question as unresolved. Request more detail rather than inventing an explanation for how the system works.

A reproducible backtest can still be limited to its assets, dates and execution assumptions. An uncertainty interval can describe a reported metric without resolving every research choice that preceded it.

Where does Anny Trade fit?

Anny Trade uses large language models for Ask Anny, its assistant. Its public Optimizer page says the indicators, backtests, parameter search and validation are “deterministic code — no LLM anywhere in that loop.” The CFO Anny Line and Optimizer math use no LLM.

Anny's bots run explicit, rule-based entry and exit rules on exchange accounts you connect, and Anny does not support grid bots. These descriptions identify where language models are used; they do not establish profitability.

Explore the Optimizer's deterministic backtests and parameter search and apply the same checklist to it.

FAQ

Can an AI trading bot predict crypto prices?

A trained model can estimate a defined market outcome, but the existence of a prediction does not establish useful predictive skill. Our task was direction over a specified horizon, not an exact future price. No reliable out-of-sample directional skill was shown by any model in this article.

Is hit-rate strategy scoring machine learning?

Performance-based strategy selection can count as statistical learning in a broad sense because it uses historical outcomes to score available strategies. The label does not tell you whether those choices will be useful later. A hit rate is not a profit measure.

Does using an LLM mean a bot uses machine learning to trade?

An LLM can write a strategy or call trading tools without being a market prediction model. Ask what it produces, which tools it can access and how execution is controlled. Then assess the resulting strategy separately; language-model use alone does not establish trading performance.

What is walk-forward validation?

Walk-forward evaluation trains on earlier data and tests on later data, advancing through time. Our procedure used a rolling training window and repeated refits, with a purge before each test block. It differs from blocked folds that train on both sides of a test block.

Can I verify a closed-source bot's AI myself?

Public documentation can help identify the claimed mechanism without resolving server-side implementation. Missing disclosure is not evidence either way. Our simplified experiments do not inspect closed-source systems and cannot establish what any vendor's undisclosed model does.

How we ran these tests (methodology and limits).

Every result in this article comes from simulations and backtests on historical data, not from actual trading. Hypothetical and simulated results have inherent limitations: they are built with hindsight, do not reflect real execution, and do not predict future performance.

The dataset described above contained 42,359 hourly bars per asset, with one exchange-side missing hour at 2023-03-24 13:00 UTC. Sampled monthly checksums matched Binance's published checksums. Software versions were numpy 2.5.4, pandas 3.1.1 and scikit-learn 1.9.1.

The data end at 2026-10-31 23:00 UTC, and incomplete outcomes were handled as follows.

• Prediction tests: bars without a complete next-24-hour label were dropped, so the last labelled 4-hour bar is 2026-10-30 20:00 UTC. The last of the 35 walk-forward blocks starts on 2026-10-17.
• Grid: a decision was kept only if 30 full days of later data existed. That gives 195 weekly decisions, from 2023-01-02 to 2026-09-21.
• Selector: it made 199 weekly decisions, from 2023-01-02 to 2026-10-19. The last three, from 2026-10-05, have 28-day windows shortened by the data end, and signals in the final 24 hours count as non-hits. Dropping those three decisions leaves the small differences in the same direction (BTC 0.32 vs 0.34, ETH 0.47 vs 0.46).

Logistic regression used standardised features with regularisation C = 0.1. The split comparison used a random forest with 200 trees and minimum leaf size 5. The non-leaky inputs covered returns over 1, 3, 6, 18 and 42 bars; volatility over 18 and 42 bars; RSI(14); the gap between 20-bar and 50-bar exponential moving averages; and a 42-bar volume z-score.

The walk-forward AUC intervals used a moving-block bootstrap with 42 consecutive 4-hour bars per block, 500 resamples and the 2.5th–97.5th percentiles. The trading illustration earned the next 4-hour return from the signal bar's close, without latency. Its 15 bps cost comprised an assumed 10 bps fee and 5 bps slippage per position change.

Although the forecast targets next-24-hour direction, this trading policy revisits its position every 4-hour bar. Its returns evaluate that particular policy, rather than direct monetisation of the forecast horizon. Neither AUC near 0.5 nor accuracy below “always up” alone rules out profitable opportunities: payoff magnitudes, thresholds and conditional performance also matter.

Grid candidates combined range widths of 0.8×, 1.0×, 1.25× and 1.6× the trailing half-range with 10, 25, 50 and 100 levels. Every fill paid 0.05%, without slippage. Each week started fresh; unsold coin was valued at the final close without an exit fee. Gap intervals use a circular moving-block bootstrap of the 195 weekly gaps (12-week blocks, 10,000 resamples, fixed seed). Weekly gaps were mildly negatively autocorrelated at lag 1, and an automatic block-length rule suggested about 11 weeks (BTC) and 16 weeks (ETH). Intervals for blocks of 2–16 weeks were similar to, or narrower than, an ordinary bootstrap's.

The selector's rules used RSI, MACD, Bollinger bands, an EMA crossover and a previous-48-hour-high breakout. A rule needed at least 5 completed signals to receive a score. Its overlapping evaluation windows matter when interpreting the small observed differences.

A second-pass internal review by the research desk reran all four original scripts and obtained byte-identical output files. At 40 of 40 randomly selected timestamps, replacing future data left non-leaky features unchanged; the deliberate look-ahead feature failed 40 of 40 checks by design. There were 0 walk-forward folds with label overlap into the test period, and the scaler was fitted on training data only inside the pipeline.

That review supports reproducibility and the absence of known leaks in the non-leaky pipelines. It does not show the results would hold in other periods, on other assets or venues, or live, and it does not remove research-selection bias.

The limits remain material: 2 assets, 1 venue, one period and simple models. Labels overlap; there is no order-book slippage or multiple-testing correction beyond what is reported. The grid simulator uses a 4-point intrabar path, on which absolute returns depend. We did not test any vendor's actual system, run FreqAI or trade live.

See Also

• What is walk-forward validation?
• How does the CFO Anny Line read market regimes?
• How do I connect my AI to live crypto market data?
• Freqtrade's lookahead-analysis documentation

Every result in this article comes from simulations and backtests on historical data, not from actual trading. Hypothetical and simulated results have inherent limitations: they are built with hindsight, do not reflect real execution, and do not predict future performance.

This article is educational only — not financial advice. Past performance does not indicate future results. Crypto assets are volatile and you can lose your entire investment. Anny Trade is an analytics platform, not a registered investment adviser.