Define R Squared (and what it is not)
R Squared (often written as R²) is a statistic that measures how much of the variation in an observed outcome is explained by a specified model over a given dataset. In plain terms, it summarizes how closely the model’s fitted line or curve matches past data.
A key limitation follows immediately: R Squared is about describing fit to data, not about guaranteeing results. A high R Squared means the chosen relationship matched the historical sample well; it does not mean the same relationship will hold later.
How R Squared works in a practical setup
To calculate R Squared, you typically compare model-predicted values to actual values for the same historical period. The statistic depends on:
- The model form you choose (for example, linear vs. nonlinear relationships).
- The dependent variable you model (for example, returns vs. price levels).
- The independent variable(s) you include.
- The dataset window and sampling choices (the specific time range and how observations are spaced).
Because these choices are part of the measurement, R Squared is not a universal property of “the market.” It is a property of the specific relationship measured on a specific historical sample.
Evidence and examples of failure modes
A common failure mode is overfitting. If a model is flexible enough to follow noise, it can achieve a strong historical fit, which inflates R Squared, even if the relationship is not stable.
Another failure mode is regime change. Markets can transition between different volatility or liquidity conditions. When the data-generating process changes, the historical relationship that produced a high R Squared may weaken or disappear, lowering future explanatory power.
A third failure mode is variable quality and changing conditions in the data pipeline. If your realized observations include effects not present in the modeling (such as transaction costs, different execution timing, or inconsistent data sources), the historical fit may not translate to realized outcomes. Even if the statistical fit looks good, the “real world” measurement can diverge.
Finally, R Squared can be misleading when the modeling target is not appropriate. For instance, modeling absolute price levels often introduces non-stationary behavior. R Squared may look high due to shared trends, while the relationship may not be meaningful for the underlying question you care about.
Limitations and risks you should verify independently
R Squared has material limitations that are worth checking explicitly:
-
It is conditional on assumptions. If the dataset window, variables, or model form change, R Squared can change too. That means you should treat R Squared as conditional evidence, not a standalone conclusion.
-
Historical relationships do not establish future results. Even a stable-looking R Squared on one period does not rule out that the relationship will break under new market conditions.
-
Costs and execution can invalidate “explained variation.” A statistical fit can ignore practical frictions, so you should distinguish between explanatory power and actual realized performance.
-
Single-number summaries can hide structure. Two models can yield similar R Squared but differ in residual behavior (for example, whether errors are random or clustered). That matters because patterns in residuals suggest the model is missing important factors.
-
Uncertainty is not directly expressed by R Squared alone. R Squared summarizes fit, but it does not automatically tell you how stable the fit is across resamples or how sensitive it is to small data changes.
Verification and next questions to make the concept testable
To independently verify what R Squared says in your use case, you can test the stability of the measured relationship rather than relying on the single value:
- Compare results across multiple time windows (e.g., earlier vs. later periods) using the same model setup.
- Check whether the relationship remains after changing the sampling frequency or time horizon.
- Examine whether residuals show systematic patterns, not just whether the overall fit looks strong.
- Confirm that the dependent variable and model form match the decision-relevant quantity you intend to explain.
A useful next question is not only “What is the R Squared?” but also “What assumptions, data choices, and market conditions are required for the relationship to stay explanatory?”