Should an AI's trading picks have a public track record?
An AI that publishes trading picks without a scored record is offering confident wording, not evidence. A real record means every pick is logged before the outcome is known, scored against actual prices at a fixed time horizon, and kept intact — hits and misses both. Without those three conditions, you are reading a curated highlight reel.
What separates a scored record from a highlight reel?
A highlight reel is assembled after prices have moved. The picks that worked are shown; the ones that did not are quietly absent. Because the selection happens retrospectively, the record looks impressive without containing any real predictive information. This is not always deliberate — it can happen simply because nobody built a system to log picks before the outcome was visible.
A scored record has three properties that a highlight reel cannot fake. First, the pick is logged at the moment it is made, with a timestamp and a target horizon — say, seven days or thirty days — so the evaluation date is fixed in advance. Second, every pick in that cohort is scored at the horizon, regardless of what happened. Third, the score is calculated from actual market prices, not from a model's own assessment of how it did. When all three conditions hold, the hit rate is a measured fact rather than an assertion.
- Timestamp
- The exact moment a pick is published, before any price movement is known. Without this, a pick can be backdated to a favourable entry.
- Fixed horizon
- A pre-declared date or time at which the pick is evaluated — for example, closing price seven days after publication. Changing the horizon after the fact is equivalent to deleting the miss.
- Cohort scoring
- Every pick made in a given period is scored together. Cherry-picking which picks count in the denominator is the most common way a real hit rate is inflated.
- Realised price
- The actual market price at the horizon, taken from a verifiable feed, not from the model's own estimate of where the asset should have gone.
Why does the time horizon matter so much?
A pick that says 'this asset will rise' is not falsifiable without a horizon. If prices fall for three months and then recover, a flexible horizon lets the publisher point to the recovery and claim the pick was correct. A fixed horizon removes that flexibility: the price on day seven is the price on day seven, whatever happens on day eight.
Different horizons also measure different things. A one-day horizon tests short-term momentum signals. A thirty-day horizon tests something closer to a thesis about fundamentals or sentiment. Mixing horizons in a single record — or switching horizons between picks — makes the aggregate hit rate meaningless, because you are no longer measuring a consistent thing.
| Horizon | What it tests | Common distortion risk |
|---|---|---|
| 1 day | Short-term price momentum | Survivorship: only volatile, high-volume assets tend to move enough to register a clear hit |
| 7 days | Near-term sentiment or catalyst follow-through | Horizon-switching: a miss at seven days is re-evaluated at fourteen |
| 30 days | Medium-term thesis or trend | Selective logging: picks are added to the record only when the thesis is already working |
| Variable | Nothing consistent | The entire record becomes unfalsifiable — any outcome can be claimed as a hit |
When is a simpler approach the better choice?
If you already have a clear, rules-based strategy — buy when a specific indicator crosses a threshold, sell when it reverses — you do not need an AI to generate picks. A simple backtesting script on historical data will tell you more about that strategy's behaviour than any AI commentary will. AI-generated picks add value when the signal is harder to formalise: synthesising news sentiment, identifying pattern combinations across many instruments, or adapting to changing market regimes.
It is also worth being honest about what a track record can and cannot tell you. Even a genuine, well-scored record is a sample from a specific market period. A strategy that performed well during a trending market may behave very differently in a mean-reverting one. A public record is evidence, not a guarantee, and anyone presenting it as the latter is overstating what the data supports.
How can you verify a record you did not watch being built?
The most practical check is to look for the picks that did not work. A genuine record will have a visible miss rate. If every published pick appears to have been correct, the record is almost certainly a highlight reel. Ask what the denominator is: how many picks were made in the period, not just how many were shown.
A second check is consistency of horizon. Read through several picks and confirm that each one was evaluated at the same declared horizon, not at whichever date produced the best result. A third check is the source of the price data. A record scored against a named, verifiable price feed — a major exchange, a regulated data provider — is harder to manipulate than one where the scoring methodology is not described.
GROX keeps a public record of its alpha picks scored against real prices at a fixed horizon, so hits and misses are counted rather than asserted. That is the minimum standard worth holding any AI pick service to.
Common questions
What is the minimum a public AI trading record should include?
At minimum: a timestamp for each pick made before the outcome is known, a declared evaluation horizon that does not change after publication, and a score derived from actual market prices. The record should include every pick in a given period — not a selection — so the hit rate reflects a real denominator rather than a curated one.
Can a good track record guarantee future performance?
No. A track record is evidence about a specific strategy during a specific market period. Market conditions change — volatility regimes, correlations and liquidity all shift — and a strategy that performed well in one environment may not in another. A genuine record is useful information, but treating it as a guarantee is a misreading of what historical data can tell you.
How do I tell if an AI's picks are being backdated?
Look for verifiable timestamps on the original publication — a post time on a public platform, a blockchain record, or a dated entry in a system you can inspect. If picks are only ever shown as a summary table with no original source you can check, backdating cannot be ruled out. A credible record points you to the original pick, not just the outcome.
Is a high hit rate always a sign of a good pick service?
Not necessarily. A high hit rate on very small predicted moves may produce worse real-world results than a lower hit rate on larger moves, once trading costs are considered. Hit rate alone does not capture the size of wins versus losses. A complete record should let you see both the frequency and the magnitude of outcomes, not just whether the direction was correct.
If you want to see what a scored pick record looks like in practice, GROX keeps every alpha pick logged against real prices at a fixed horizon — you can read the hits and the misses yourself.