Definition and what “different behaviour” means
R Squared (R²) is a statistical measure of how much of the variation in a dependent variable is explained by an independent variable(s) in a model. In plain terms: it describes how closely the model’s past predictions match historical outcomes, using the same data that was used to fit the model. It does not inherently indicate future trading performance or the likelihood of a particular move.
So when people say R Squared “behaves differently” under certain market conditions, they usually mean: calculated R² values change because the underlying relationship between inputs and outcomes changes, and because the quality of the data and assumptions behind the model changes.
Mechanism: why market conditions can change R²
R² depends on three broad elements: (1) the relationship strength between the variables, (2) the noise level, and (3) how stable that relationship is across the sample.
-
Relationship strength (signal vs. randomness) If the market environment produces a stronger, more repeatable relationship between the variables you model, the fitted curve tends to track historical outcomes more closely, raising R². If the relationship weakens or becomes inconsistent, R² can fall.
-
Noise and structural breaks High volatility, sudden regime shifts, and changing microstructure can increase noise. When noise dominates the data, even a reasonable model may fit less closely, reducing R².
-
Sample composition and scaling R² is computed from a particular sample window and choice of variables. Changing the timeframe, the amount of historical data used, or the transformation (for example, using returns versus raw prices) can change how much variance is captured.
Evidence by example (with explicit assumptions)
Consider a simple modeling setup: you fit a regression between a chosen input series and an output series using a fixed rolling window length. Assume you evaluate two regimes on comparable sample sizes.
- Regime A: smoother movement with a more consistent input–output relationship. If the model error is smaller on many observations, the explained variance rises and R² tends to be higher.
- Regime B: frequent switching between behaviours (for example, alternating trending and mean-reverting phases). If the input–output relationship changes over time, the same model form may fit some parts of the window but fail in others, increasing residuals and often lowering R².
This is not proof that any regime will “always” produce higher or lower R². It only illustrates the mechanism: R² is sensitive to stability and noise.
Limitations and failure modes
Key limitations are worth separating from “market conditions”:
- Overfitting to the sample: A model can fit historical data well even if the relationship is temporary. A high in-sample R² can be misleading about out-of-sample behaviour.
- Regime dependence: A market regime can change after you compute R². Historical relationships do not guarantee future results.
- Variable mismatch: If the independent and dependent variables are not aligned in time (for instance, using inputs that do not correspond to the output window), the computed R² may change dramatically.
- Cost and execution effects: R² computed on price series ignores trading costs, slippage, and execution constraints. Even if the fitted relationship is strong statistically, real-world outcomes can diverge.
- Nonlinearity and model form: R² is tied to the chosen model structure. If the true relationship is nonlinear or changes form, R² may understate usefulness or fluctuate.
Verification: how to independently check your assumptions
To verify “when R² changes,” you can check how it varies across: (1) different sample windows, (2) different timeframes, and (3) different model specifications. Record whether R² changes primarily because the data become noisier, because the relationship becomes less stable, or because your transformation choices change variance.
A next question to consider is: are you computing R² on the same type of series (e.g., returns versus prices), with consistent alignment and sample length? If not, “different behaviour” may come from the calculation setup rather than from the market itself.