What is per-ticker backtest CI90 accuracy?
Backtest accuracy here means per-ticker CI90 coverage: how often this specific ticker's historical 90% confidence bands actually contained the realized price, measured by rewinding the model and checking its past forecasts against what really happened. A reading near 90% is well-calibrated for this name; well below 80% means the model has historically been overconfident on this ticker — its bands were too narrow for how much the price actually moved.
Per-ticker vs. platform-wide calibration
This is deliberately a per-ticker number, distinct from the platform-wide calibration page, which aggregates coverage across the committed TOP-20 backtest by market cap. The aggregate figure can look well-calibrated (85-95% range) while individual names remain poorly covered — structural-break tickers like a company mid-acquisition or undergoing a business-model shift are the classic case where per-ticker coverage lags the aggregate badly, even at a 1-year horizon. Always check the per-ticker number for a name you're actually trading, not just the platform average.
Live example: AAPL's own 3-month backtest CI90 coverage is 98.9% — of past 90% bands the model produced for this ticker, this is the share the realized price actually fell inside. See the full AAPL forecast for the current band this backtest is calibrating.
This is coverage, not a directional win rate
Backtest CI90 accuracy says how often the band contained the outcome — it is not the same as how often a BUY signal on this ticker would have been directionally correct or profitable. Directional accuracy and win rate are separate, harder-to-move diagnostics; conflating the two is one of the most common ways backtest numbers get over-read.
Frequently asked questions
Why can a ticker have worse backtest accuracy than the platform average?
The platform-wide calibration page aggregates coverage across the TOP-20 backtest set; individual names, especially those going through a structural business change, can have materially worse per-ticker coverage than that aggregate.
Is backtest CI90 accuracy the same as a win rate?
No. It measures how often the confidence band contained the realized price — a coverage statistic — not whether a directional BUY/SELL call on the ticker was correct or profitable.
Where can I see the platform-wide version of this figure?
The live calibration page reports aggregate CI90 coverage per horizon across the committed TOP-20 backtest, including the misses alongside the hits.
Educational research only — not investment advice.