Evaluate AI crypto signals by checking the evidence behind the label. Ask what the system does, which records support its claims, and whether those records describe something you could have acted on. A polished explanation or a winning screenshot cannot answer all three.
Our guide to reading trading signals covers entry, invalidation and timing. This guide evaluates the provider and its supporting record before you consider following a setup. The checklist and example are educational tools, not a provider ranking or a recommended strategy.
Ask what AI actually does
Have the provider identify the task: summarizing information, ranking candidates, estimating an outcome, or generating a proposed trade. Ask which inputs it uses, how fresh they are, whether a person reviews the output, and which model or strategy version produced the signal. An automated indicator rule is not automatically AI; AI and technical analysis can also be combined.
Trace an explanation back to its underlying sources. A fluent summary is not verification of a price, announcement or event time. The joint SEC, NASAA and FINRA investor article warns that AI-generated information can be inaccurate, outdated or fabricated, even when its inputs are accurate. Check the provider’s identity through independent sources rather than trusting a profile or testimonial alone.
Separate three kinds of record
A backtest simulates past operation under stated assumptions. A forward paper record logs signals as new information arrives but still simulates execution. An executed account record documents actual fills and account activity. Ask which one a chart shows; “live” by itself does not settle the question.
Request the complete reporting period, markets, strategy versions, trade list and treatment of open positions. For a backtest, ask when training and strategy selection ended and which later data were kept separate for evaluation. Repeatedly adjusting a model after seeing that evaluation weakens its independence. A large trade count concentrated in one market regime leaves other conditions untested.
Use the backtests and live results guide for definitions of return, drawdown and sample coverage. Here, the practical question is whether the provider supplies enough records to inspect those measurements.
Use an evidence checklist
Save the following worksheet with a review date. For each row, record the source document or export, what you verified, and what remains unknown. “Not supplied” is a finding, not a zero.
| Check | What to request and verify |
|---|---|
| Provenance | Provider identity, input sources, timestamps, model or strategy version, and the role of human review. |
| Complete record | All signals in a defined period, including losses, cancellations, untriggered ideas, open positions and revisions. |
| Evaluation boundary | Training and selection dates, untouched evaluation data, market coverage and separate results for changed versions. |
| Execution and costs | Publication and receipt times, reference prices versus fills, fees, spread, slippage and applicable holding costs. |
| Limits and incentives | Missing data, outages, losing periods, access charges, referral payments and any conflict affecting the presentation. |
A provider need not reveal proprietary source code for you to ask these questions. If essential evidence is unavailable, you cannot verify the related claim. An independent audit can help only to the extent that its scope, period and method cover that claim.
Reconcile a hypothetical provider report
Imagine a fictional provider publishes 20 signals during a fixed observation window. Its complete log identifies 12 closed paper trades, 3 still-open paper trades, and 5 ideas that never triggered. None represents an actual exchange fill. The counts reconcile: 12 + 3 + 5 = 20.
Of the closed paper trades, 8 made a gross gain and 4 a gross loss. Thus 8 ÷ 12 ≈ 66.7% describes the gross win rate among closed paper trades. It says nothing about the eventual outcome of the three open positions or your next trade. Counting 8 ÷ 20 = 40% answers a different question: the share of all published ideas already closed with a gross gain. Neither percentage establishes profitability.
Suppose the stated simulated sizes produce 24 USD aggregate gross profit across those 12 closed trades. If the assumed total trading costs for the same trades are 30 USD, their net result is 24 − 30 = −6 USD. This excludes the three open positions and subscription charges; it is not the complete portfolio result or a percentage return. Every count and amount here is invented to demonstrate reconciliation, not observed performance or a suggested fee level.
Check what a follower could have received
A timestamped publication record is stronger evidence of availability than a screenshot reconstructed later. Compare publication, receipt and earliest possible action. A model’s reference price may have passed before a subscriber received the signal. Missing or edited messages need an explanation and a preserved revision history; a later chart should not silently replace the original proposal.
Request cost assumptions appropriate to the instrument and venue. Avoid double-counting spread or slippage already included in simulated fill prices. Keep subscription charges separate and show their effect over the same evaluation period. The CFTC’s AI trading advisory cautions against guaranteed-return claims and highlights fees, spreads and subscription costs. AI branding does not remove execution risk.
Finish with a documented decision
Write one of three conclusions: continue researching, with the next check; wait for evidence, naming the missing item; or reject the claim, stating the contradiction or unsupported promise. None means a trade has been approved.
Apply the same standard to every provider. Keep the worksheet, source links and dated records so a later model change or marketing update cannot rewrite your original review. Educational information, not financial advice; crypto trading can involve substantial losses, and historical or paper results do not guarantee future outcomes.
