FA

Forecast Arena

Compare models and see when they win.

HomeHelp

Help

How to use the app, what each section means, and how to prepare the input file.

1. Upload your dataProvide an XLSX with reality, predictions, and optional exogenous columns.
2. Review the resultsSee metric summaries, the recommended model, and side-by-side charts.
3. Act on insightsUse the recommendations for future forecasts and track the most influential segments.

What this app is for

Forecast Arena helps reveal which model performs best in different situations and how stable those results are.

  • Compare two or more forecasting models on your data.
  • Identify when and why each model is more accurate.
  • Receive recommendations for future predictions even when ground truth is missing.

How to use the app

  1. On the landing page, upload an Excel file (XLSX) with the required columns.
  2. After processing, explore the summary, recommendation, charts, and segments.
  3. Use zoom/pan in charts, filter with the legend, and try the custom prediction feature.

If you install the app as a PWA, you can keep using it offline.

Input file (XLSX)

The app reads the first sheet in the workbook. Make sure required columns and names are present, otherwise the import fails.

Required columns

  • ds — date/time (ISO or Excel date code).
  • reality — ground truth value (may be empty).
  • prediction1, prediction2 — numeric predictions.

Optional columns

  • prediction3, prediction4… — additional models to compare.
  • Any other columns are treated as exogenous features (numeric or categorical).

Example (CSV for illustration)

ds,reality,prediction1,prediction2,weather,temp
2024-01-01T00:00:00Z,120,110,130,sunny,2.3
2024-01-01T01:00:00Z,125,121,134,cloudy,2.0
2024-01-01T02:00:00Z,,118,136,cloudy,1.8
Download sample XLSX

Rows without reality feed the recommendation, while metrics use only rows with reality.

Sections explained

Result

Summary metrics, the long-term MAE winner, and key differences.

Recommendation for missing reality

Recommended model, custom prediction (KNN), and closest historical matches.

Predictions vs. reality

Compare how the prediction lines follow reality (filled area). Look for: (1) long‑term bias (above/below), (2) lag or overshooting during fast moves, and (3) accuracy in extremes. Zoom into periods and see which model better tracks local peaks and troughs — that hints which model to choose in similar conditions.

Error differences over time

See when — and by how much — each model is more accurate. Higher lines mean higher absolute error; the filled area shows the error difference (Model 1 − Model 2). Long red/green runs indicate consistent superiority; larger filled area means a bigger margin. Zoom to find intervals with large gaps — choosing the right model matters most there.

Effect of exogenous variables on error

Values describe the strength of relationship between a variable and absolute error of each model. Higher means stronger effect. Positive increases error; negative reduces it.

Segments and trend

Segments show exogenous influence; the trend summarises the latest period.

Chart tips

  • Zoom: drag or mouse wheel, pan: Shift + drag
  • Use the legend to show or hide series as needed.
  • Numbers and axes follow the selected language and formatting.

Tip: Click and drag on a chart to zoom; double-click to reset the full period.