What is MAPE in a stock forecast accuracy check?
MAPE — Mean Absolute Percentage Error — measures how far off the model's point (mean) forecast has historically been from what actually happened, as a percentage of the realized price. It is computed on the model's own out-of-sample backtest for each ticker: rewind the model to an earlier date, forecast forward, and compare the mean forecast to what the price actually did. Lower MAPE means a more accurate point forecast. Roughly: under 10% is strong, 10–25% is workable, and above 25% signals a genuinely hard-to-forecast name.
MAPE is not the same as calibration
MAPE grades the single point estimate; CI90 coverage grades the band around it. A ticker can have low MAPE (the mean forecast is usually close) but poor CI90 coverage (the stated 90% band is too narrow to actually contain 90% of outcomes) — these are separate diagnostics and both matter. A model that is accurate on average but overconfident about its own uncertainty is arguably worse to trade off than one with wider, honestly-calibrated bands.
Live example: AAPL's 3-month MAPE from the last backtest run is 7.4% — the average absolute error of the model's point forecast versus the realized price on this ticker's own history. See the full AAPL forecast.
Why MAPE varies so much by ticker
MAPE is not a single platform-wide number — it is measured per ticker and per horizon, because forecastability genuinely differs across names. A stable, low-beta consumer staple typically backtests to a lower MAPE than a high-beta, high-volatility growth name or a stock mid-way through a structural business change. Quantustik publishes MAPE per ticker rather than a single blended average so this dispersion is visible, not hidden.
Frequently asked questions
What counts as a good MAPE for a stock forecast?
Roughly: under 10% is strong, 10-25% is workable, and above 25% signals a genuinely hard-to-forecast name. MAPE varies a lot by ticker, so it's read per name, not against one universal bar.
Is low MAPE the same as good calibration?
No. MAPE grades the accuracy of the single point forecast; CI90 coverage grades whether the confidence band's stated 90% actually holds. A model can be accurate on average yet overconfident about its own uncertainty.
How is MAPE computed?
By rewinding the model to an earlier date, generating a forecast, and comparing the mean forecast to the price that actually happened — repeated across the historical backtest window for each ticker and horizon.
Educational research only — not investment advice.