Direct answer
“Divergence in R Squared” means the explained-fit value (R Squared) does not stay consistent when you compare it across different samples, time windows, datasets, or calculation settings. In plain terms, the relationship that produced the earlier “fit” becomes weaker, stronger, or qualitatively different later. This can happen even when the same formula is used, because real relationships in data often change over time.
Mechanism or definition
R Squared is a statistic that summarizes how well a model’s outputs track a target variable in a particular dataset. A typical intuition is: if R Squared is higher, the model’s predictions align more closely with the observed values in that dataset.
When someone talks about “divergence,” they usually mean one of these comparisons:
- Across time windows: R Squared calculated on recent data differs from R Squared calculated on older data.
- Across assets or series: R Squared differs when repeating the same model on different instruments or groups.
- Across recalculations: R Squared changes when the underlying inputs (for example, which observations are included) change.
To make this concrete, consider a simple example with explicit assumptions: you compute R Squared from a linear regression of a response variable on a feature, using observations from two separate periods (Period A and Period B). If Period A has higher R Squared than Period B, the fit “diverged.” The calculation itself is not wrong; the data-generating relationship is simply different between periods.
Evidence or example
A common failure mode is treating “fit” as stable. Suppose you calculate R Squared on a training window and then later check it on a newer window. If the newer-window R Squared is much lower, that divergence suggests the model is not capturing a stable relationship.
Another example: you might test the same idea with two different feature sets. The first setup could show strong R Squared because it captures noise patterns in the specific sample. The second setup might generalize less well, giving a different R Squared. That divergence is a clue that the earlier result depends on the specific dataset or assumptions.
In market-related contexts, especially where conditions shift, divergence often indicates non-stationarity: the statistical relationship between inputs and outputs changes over time. However, it is important to separate two meanings:
- Stable mechanics: the R Squared formula and regression setup are fixed.
- Variable environment: the data relationship changes due to changing dynamics.
Limitations and risks
R Squared divergence is not a standalone trading indicator or a promise of outcomes. Key limitations include:
-
It can reflect non-stationarity, not “direction.” Divergence only tells you that the fit changed. It does not tell you whether future results will improve or deteriorate.
-
Hindsight bias can distort interpretation. If you look at R Squared first, then choose windows, parameters, or evaluation periods to make the result look convincing, you may overstate what is “meaningful.” The divergence you notice might be influenced by how you selected the comparison.
-
Different assumptions change the result. R Squared depends on the model form and how inputs are prepared. If two teams compute it with different preprocessing, window lengths, or definitions of the target, “divergence” may come from methodology rather than from a real underlying change.
-
Historical relationships do not establish future results. Even if divergence is consistent across past samples, there is no guarantee the same pattern will occur again.
Verification or next question
To independently verify what divergence means in your specific case, you can repeat the same calculation consistently across clearly defined samples:
- Use the same model specification and the same definition of the response and predictors.
- Compute R Squared on multiple, non-overlapping time windows.
- Report both the magnitude of R Squared and whether the change is consistent across windows.
A useful next question is: What exact samples and settings produced the divergence you are seeing? If you can specify the time windows (or datasets), the model definition, and the calculation steps, you can distinguish real relationship change from measurement or selection effects.