Horizon learning guide · English
How to Evaluate AI Trading Forecast Accuracy
Learn to assess AI market forecasts using sample size, candle timing, comparison baselines, no-call counts and price-range coverage.
A percentage needs a definition
An accuracy percentage is useful only when you know what was predicted, when it was predicted and how it was scored. Correct direction, close-price error and range coverage answer different questions. None of them, alone, demonstrates that a trading strategy would be profitable. Horizon’s public record is a place to inspect measured forecast outcomes, not a guarantee of future performance.
Compare forecasts at the same stage of the candle
A direction call made near the end of a candle has already observed most of that candle. Comparing it with an early call without identifying timing can make the results misleading. The Horizon record separates calls into progress groups: the first 10%, between 10% and one third, between one third and two thirds, and the last third. Compare like with like before drawing a conclusion.
Read the sample size and the baseline
Check how many calls produced the reported percentage. A result based on a handful of calls is much less informative than a longer record across changing conditions. Horizon leaves out markets and checkpoints with too few calls. Beside the direction results, inspect how often the same candles closed up: that is the comparison for always guessing up. A high hit rate during a rising market needs this context.
Keep no-call readings visible
A system can improve its reported directional hit rate by making fewer calls. This may be sensible, but readers need to see the coverage. Horizon reports readings where it did not make a direction call. Its public scoring definition treats readings above 65% as bullish and below 35% as bearish, with no call in between. Use the live methodology on the record page as the current reference.
Score the range separately
Range coverage asks whether the finished candle’s close fell inside the forecast range. It does not say that price stayed inside the range throughout the candle. The public record describes a range designed to contain the close about 80% of the time; this is a design aim, not a guaranteed observed success rate. Interpret coverage together with range width, market and timeframe.
Include misses and unresolved outcomes
A useful journal preserves the original reading and records its outcome consistently. Keep wrong calls visible. Note any readings that cannot yet be scored, rather than quietly treating them as correct. On the public record, inspect the latest calls as well as aggregate tables. The reported window covers forecasts run on the site, so heavily used markets contribute more observations.
Do not confuse forecast accuracy with returns
A correct direction can coincide with a move too small to cover fees or slippage. Trading results also depend on entries, exits, position size and risk. A direction score does not measure those choices. Use forecast records to judge the particular forecasting question they answer. A separate strategy evaluation would be needed to assess an execution rule.
A checklist for reading the record
Before sharing a result, identify the market, timeframe, reporting window, stage of the candle, number of calls, comparison baseline and no-call count. Quote the context alongside any percentage. Visit the live Horizon track record rather than relying on a number copied into an older article.