Regression fit is not forecast evidence

The question answered by a fitted line

Ordinary least squares chooses coefficients that minimize squared residuals on the observations supplied. For a line with an intercept, the fitted response is a + b times x. The slope is the centered cross-product sum divided by the centered square sum of x; the intercept makes the line pass through the two sample means. These are descriptions of the fitted sample, not statements about when its inputs became observable. NIST's least-squares treatment establishes this estimation problem. A regression using a contemporaneous market return can describe exposure without supplying information that was available before that return occurred.

Availability defines a forecast

A forecast has an origin, an information set and a target horizon. An input measured after the forecast origin cannot support a forecast at that origin, even when it has an excellent statistical relationship with the target. Revised economic releases, closing prices used before their close and full-sample normalization can all cross that boundary. Scikit-learn's leakage documentation distinguishes applying an already fitted transformation from learning that transformation using test observations. The same distinction applies outside machine learning. A timestamped design identifies the response interval, the latest permissible predictor observation and the point at which coefficients were frozen.

What fitted R-squared measures

With an intercept and a nonconstant response, fitted R-squared is one minus residual sum of squares divided by the centered total sum of squares. It compares the fitted line with a constant equal to the training response mean on those same rows. Adding predictors cannot reduce ordinary training R-squared when the original model remains nested in the larger model. That mechanical improvement is not new predictive evidence. Adjusted R-squared penalizes parameter count but still evaluates the fitted sample. Neither measure identifies causation, establishes economic usefulness or substitutes for a comparison against an independently specified forecasting baseline.

A high fit followed by a miss

The table is an original synthetic hypothetical with decimal returns. Its first five rows give a = 0.005 and b = 1.4. Their residual sum of squares is 0.000040 and centered total sum of squares is 0.002000, producing R-squared of 0.98. Assume the final row's x is known at a later forecast origin and its y becomes known afterward. The frozen line predicts 0.047 against a constructed outcome of 0.002. Its squared error is 0.002025; the frozen training-mean forecast of 0.005 has squared error 0.000009. This deliberately constructed contrast illustrates different questions, not an observed market result.

A point estimate leaves uncertainty unresolved

Uncertainty about the conditional mean differs from uncertainty about an individual future response. NIST's prediction-interval discussion adds future observation noise to uncertainty in the estimated mean, making the prediction interval wider under the same classical model. Exact textbook intervals rely on assumptions such as independent normal errors with constant variance and a correctly specified conditional mean. Financial residuals need not satisfy them. Uncertainty in a future predictor adds another issue when x itself must be forecast. A calculator returning coefficients and fitted values has not thereby computed either interval, tested its coverage or incorporated uncertainty in future inputs.

Extrapolation and changing relationships

The final synthetic predictor lies beyond the training range. Extrapolation asks the line to retain its form where the fit supplied no observations; a straight extension on a chart does not supply that evidence. Predictions inside the range can also fail if the relationship changes or important variables were omitted. Residual patterns, influential observations and instability across windows help describe those limitations but do not automatically repair them. Two distinct concerns remain: uncertainty about coefficients within a specified model and uncertainty about whether that model is an adequate description. Small reported coefficient errors address only the former under their assumptions.

Evidence after the fit is frozen

An empirical forecast comparison records losses on observations excluded from coefficient estimation and model selection, using the same target and information cutoff for the model and baseline. Repeated examination of those losses can turn an apparent test set into development data. Dependence among forecast errors also affects how much independent evidence a long table contains. A single synthetic miss neither invalidates regression generally nor measures expected forecast loss. The defensible conclusion here is narrower: the arithmetic reconstructs a fitted relationship, while predictive value remains a separate empirical claim requiring untouched observations, a declared comparison and an uncertainty analysis.

Primary references

These sources support methods and conventions. The example values are original constructions, not observations extracted from these documents.

• NIST e-Handbook: Least Squares
• NIST e-Handbook: Predicting a single response and its uncertainty
• Scikit-learn documentation: Common pitfalls and data leakage

Related investigations