Forecast degradation for longer horizons despite homogeneous input data

[BelhsanHmida]

Problem

When forecasting over longer horizons (e.g. 48h) using limited training data (e.g. 2 days), the forecast quality degrades significantly in the later part of the horizon.

In a controlled test with nearly identical daily patterns:

• The first 24h of the forecast is reasonably accurate
• The second 24h shows strong degradation and drift

Given the homogeneity of the data, some degradation is expected, but not to this extent.

Example:

Why this matters

This behavior can be confusing for users, especially when working with clean or repeating signals where a stable forecast is expected across the full horizon.

It may indicate that:

• The model configuration is not well-suited for longer horizons
• Seasonality (e.g. daily patterns) is not properly captured
• The forecasting setup becomes unstable as the horizon increases

Steps to reproduce

1. Use a dataset with a clear and repeating daily pattern (e.g. 3 identical days)
2. Train the model on the first 2 days
3. Forecast the next 48 hours
4. Compare forecast vs actual:
   • First 24h → relatively accurate
   • Next 24h → significant degradation

Current behavior

Forecast quality drops significantly in the second half of the prediction horizon, even when the input data is highly regular.

Next steps

We will investigate this behavior further to better understand its root cause and determine whether improvements in model configuration, horizon handling, or seasonality capture are needed.

Possible directions

• Investigate how model configuration affects longer horizons (e.g. lag settings, window sizes)
• Ensure daily seasonality is properly captured or configurable
• Evaluate whether different forecasting strategies behave more robustly for multi-day horizons
• Add validation or warnings when training data is insufficient for the requested horizon

[nhoening]

dummy-data-3days.xlsx

The data set for 3 days which I used when I encountered this.