Why the previous-award baseline matters
A model can improve substantially on a weak comparison and still add very little to a strong one. That distinction matters in my utility returns study.
The comparison
I fit the models on 345 observations from 2010-2019 and evaluate them on 188 later observations from 2020-2024. The market-and-case model has a mean absolute error of 53.1 basis points. Adding the previous award lowers that to 23.3.
Those two numbers look like a large gain. But using the previous award itself, without estimating a new regression, gives 24.2 basis points. The augmented model improves on that baseline by 0.9 basis points.
| Prediction rule | MAE, basis points |
|---|---|
| Training-period mean | 42.2 |
| Previous award | 24.2 |
| Market and case variables | 53.1 |
| Model including previous award | 23.3 |
What this changes
The previous award contains substantial predictive information. That does not show whether commissions consciously anchor on it. Stable utility characteristics, regulatory mechanisms and commission policies can also produce persistence.
For prediction, the next question is whether extra variables improve on the previous-award rule consistently across settings. For explanation, the next question is which mechanisms account for the persistence. They require different evidence.
Timing matters
The regression coefficients remain fixed after the training period. An earlier decision in the test period may inform a later one. The economic data contain revisions made after the decisions, so this is a retrospective evaluation rather than a reconstruction of information available in real time.
I report the simple rule alongside the regressions so a reader can judge what the additional complexity contributes.