Definition
Temporal bias occurs when machine learning models disproportionately weight recent observations compared to older ones, often due to non-stationary data distributions or specific training protocols. This can result in models failing to generalize across time, missing long-term trends, or exhibiting drift as the underlying data patterns evolve. It is critical in time-series forecasting and dynamic systems to mitigate this bias to ensure robustness and fairness over extended periods.
Summary
A systematic error where models prioritize recent data over historical context, leading to skewed predictions.
Key Concepts
- Data drift
- Non-stationarity
- Recency effect
- Model decay
Use Cases
- Financial market prediction
- Social media trend analysis
- Churn rate modeling