Time-Series Forecast Metrics for Accuracy, Dependence, and Uncertainty

Summary

This Pine library collects statistical measures for assessing time-series forecasts. Its point-forecast metrics include absolute, squared, root mean squared, normalized root mean squared, and percentage errors. It also offers interval-based calculations that turn interval bounds into a midpoint forecast before scoring them. These measures help compare forecast errors on different scales and summarize deviations between predictions and observations.

The library also covers dependence and probabilistic forecasts, including autocorrelation, an augmented Dickey-Fuller test, interval coverage and sharpness, resolution, Theil-style inequality, and pinball loss for quantiles. The code and descriptions explain the available functions, but provide no trading strategy, benchmark dataset, or comparative results. Users must check input alignment and metric definitions: some functions have different size checks or offsets, percentage measures can behave poorly around zero denominators, and normalized error depends on the target range. It is a toolkit for evaluation rather than evidence that any model will perform well in markets.

Key ideas

• MAE, MSE, and RMSE summarize forecast errors with different sensitivities to large deviations.
• Normalized and percentage error metrics can aid comparison across scales, but their denominators affect interpretation.
• Autocorrelation and stationarity testing help examine time-series dependence and structure.
• Interval scores assess coverage and concentration, while pinball loss evaluates quantile forecasts asymmetrically.
• The library supplies calculations but does not provide market validation or model performance evidence.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.