Why Squared Residuals Matter for Forecast Variance
Summary
This exchange explains why a time-series model needs assumptions about both residuals and their squares. Independent, identically distributed residuals support inference about the conditional mean, while stable squared residuals help estimate forecast variance and construct confidence intervals or bands. The distinction matters because returns can have a stable overall distributional description yet still show time-varying volatility.
The answer uses a simple comparison: two forecasts can share the same expected value while one has uncertain variance and the other has a known variance. It argues that variance information is needed for uncertainty ranges and multi-step forecasts, not just a point prediction. The discussion is conceptual rather than a formal treatment of tests or model specification, and it compresses several assumptions into the IID framing; in practice, residual and volatility diagnostics depend on the chosen model and forecasting horizon.
Key ideas
- Residual behavior is assessed after fitting a model, rather than by examining raw returns alone.
- Residual dependence affects inference about the conditional mean forecast.
- Squared residual behavior helps determine whether forecast variance can be estimated reliably.
- A point forecast without a variance estimate does not provide a usable confidence interval.
- Multi-step forecasts and confidence bands require assumptions about how uncertainty evolves.
Tags
Full text
# Answer by Malick (score 0, accepted) # Why should we care if the "squares of returns are independently distributed over time" to choose an adequate model of the distribution of returns? In a Time Series Book by Hashem Pesaran, he mentions that there are a number of issues that need to be addressed in order to choose an adequate model for predicting asset returns. I understand the other 4 considerations in the picture but I don't understand what it means for the squares or absolute values of returns to be independently distributed over time? Why is that different from the distribution being constant over time which includes the variance being constant over time? ## Answer by Malick (score 0, accepted) https://quant.stackexchange.com/a/36598 First it is not returns but residuals and squared residuals of your model that should be IID. Second, it is required that squared residuals are IID because you are not only interested on your mean forecast but also about the variance of your forecast. Imagine two cases : - your forecasted value is 0.5 but its variance is unknown because residuals of your models have time-varying variance. The variance of your forecast may be 0.2 or 0.8, you simply don't know it. (in fact you only know its central tendency) - your forecasted value is still 0.5 but you know for sure that it has a certain variance (let's say 0.2) because it is derived of the variance of your squared residuals that is constant. In the second case you are able to build a confidence interval about your forecast but not in the first case. To sum up, residuals must be IID to be sure that the mean forecast is correct. Squared residuals must be IID to infer the variance of your forecast and to be able to build multi-step forecast and confidence bands.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.