Forecast Combination: Equal vs. Estimated Weights

A forecast combination takes two or more predictions for the same quantity and produces one forecast from them. If models A, B, and C each predict tomorrow’s realized variance, their outputs can be averaged or assigned different weights. The combined number is a new forecast, not a vote about which model is correct.

The inputs must refer to the same target, forecast origin, horizon, units, and information cutoff. A one-day variance forecast cannot be averaged directly with a monthly volatility forecast just because both are called volatility. First define a common quantity and interval, then align the predictions.

Combining forecasts is useful when several plausible models capture different features of a series, or when it is unclear which model will remain strongest. The forecasts need not come from unrelated algorithms. What matters for the combination is how their errors behave together, not how different their names sound.

Bates and Granger’s original application found that a composite forecast could have lower mean squared error than either of two component forecasts in their airline-passenger example. That result motivates testing combinations; it does not guarantee an improvement for another market, target, period, or loss function (Bates and Granger, 1969).

For \(K\) forecasts, the equal-weight combination is

\[ \widehat{y}_{t+h\mid t}^{\,EW} = \frac{1}{K}\sum_{k=1}^{K}\widehat{y}_{k,t+h\mid t} \]

Each component contributes the same share. This rule needs no training sample for its weights and is easy to reproduce. It is a useful benchmark even when the final method will estimate or constrain weights.

Equal weighting does not mean that all forecasts are equally accurate, independent, or economically important. It says only that the combination procedure does not estimate separate weights from past performance. An equal-weight forecast can still inherit a shared bias if all of its inputs miss the same feature.

The number of model names is not a measure of information diversity. Three volatility models may use the same return series, session boundary, and sampling interval, and may all miss the same overnight move. Averaging them does not automatically repair that omission. Check whether archived out-of-sample errors differ in useful ways; do not assume that different labels create independent evidence.

Suppose three models forecast next-session volatility as 12, 15, and 20 basis points. Their equal-weight forecast is \( (12+15+20)/3 = 15.67 \) basis points. This arithmetic shows how the blend is formed; one forecast origin cannot show whether the rule is accurate or profitable.

For volatility, check whether the input is variance or standard deviation. Averaging three variance forecasts and then taking a square root is not generally the same as averaging their standard-deviation forecasts. Use the same target definition and units that will later be scored.

An estimated linear combination may be written as

\[ \widehat{y}_{t+h\mid t}^{\,C} = \sum_{k=1}^{K} w_k \widehat{y}_{k,t+h\mid t} \]

One common theoretical result assumes each forecast error has mean zero, the forecast target is the same, and the weights sum to one. Let \(\Sigma_e\) be the covariance matrix of the component forecast errors. If this matrix were known and invertible, the minimum-variance weights under those restrictions would be

\[ w^{*} = \frac{\Sigma_e^{-1}\mathbf{1}}{\mathbf{1}'\Sigma_e^{-1}\mathbf{1}} \]

This formula is a reference point, not an instruction to invert a noisy sample covariance matrix. In practice, \(\Sigma_e\) is estimated from a finite archive of forecast errors. Highly correlated inputs can make the estimate nearly singular, so small data changes may produce large weight changes.

For two forecasts with equal error variance \(\sigma^2\), the variance of their equal-weight average is \(\sigma^2(1+\rho)/2\), where \(\rho\) is the correlation between their errors. If each error variance is \(4\) and their correlation is zero, the average has variance \(2\). If the correlation is \(0.90\), its variance is \(3.8\). The models’ error dependence, not their labels, determines how much variance averaging can remove. This calculation concerns error variance; if the component forecasts are biased, their mean squared errors also include squared bias.

Another approach regresses the realized target on archived component forecasts, often with an intercept. Granger and Ramanathan discuss linear regression methods in which weights need not add to one. The intercept and unconstrained weights can correct average bias in a training sample, but they change how the result should be interpreted: the coefficients are fitted forecast adjustments, not necessarily portfolio-style shares (Granger and Ramanathan, 1984).

Specify restrictions before evaluation. Nonnegative weights that sum to one create a convex average that stays between the smallest and largest component forecast. Allowing negative weights or an intercept can produce values outside that range. That may be appropriate for a calibrated forecast, but it can be unsuitable when the target must remain nonnegative, as variance does.

Assume three models issued forecasts for the same next-session target: 12, 15, and 20 basis points. Suppose a weighting rule estimated on earlier forecast origins assigns \(0.50\), \(0.30\), and \(0.20\). The resulting forecast is

\[ 0.50(12) + 0.30(15) + 0.20(20) = 14.5\ \text{basis points} \]

The weights sum to one and are nonnegative, so this is a convex combination. It falls between the component forecasts. The difference from the equal-weight result of 15.67 basis points reflects the chosen weights; it does not show that the weighted forecast is better.

The weights must have been estimated only from forecast errors whose outcomes were already known at the time of estimation. If this is a historical backtest, the 14.5-basis-point forecast at origin \(t\) may use information available by \(t\), not the realized target at \(t+1\) or later.

After the target becomes observable, record the component forecasts, combined forecast, actual proxy, and evaluation loss. Repeating that process over later forecast origins lets you compare rules on common dates. It still does not turn one worked calculation into evidence of a persistent advantage.

The best weights for known population error covariances are not the same as weights estimated from a short sample. Estimated weights add another layer of uncertainty. When component forecasts are similar, their errors can be highly correlated and the optimization may try to exploit small historical differences that will not persist.

Kang’s simulations and macroeconomic-forecast application found that combination weights could be unstable and that a simple average could perform better in practice in those settings (Kang, 1986). Timmermann’s review describes how estimation error, model instability, and error dependence can affect forecast combinations (Timmermann, 200601004-9)).

Later analysis of the forecast-combination puzzle makes the same finite-sample point: results derived for fixed “optimal” weights need not carry over unchanged when those weights are themselves estimated (Claeskens et al., 2016). These findings do not prove that equal weights always win. They explain why a more complicated weighting rule must earn its place in an out-of-sample comparison.

<!-- learn:illustration -->

Possible responses include beginning with equal weights, limiting the number of inputs, shrinking estimated weights toward equal shares, or imposing a simple constraint. Each choice trades flexibility against estimation noise. Select it using training data and report the rule; do not choose the most attractive version after looking at the final test period.

Every archived component forecast should be one that could have been produced at its stated origin. Recreate each model using only data available then, preserve the forecast timestamp, and save the model version and target definition. In-sample fitted values are not substitutes for genuine historical forecasts because the model may have used observations that were not yet known.

Align the archive by target and origin. For every row, the component forecasts must target the same future interval, and their training information must stop at the same decision time. Drop or explicitly handle missing forecasts before estimating a combination; silently changing the set of inputs can make weights mean different things on different dates.

Keep weight estimation separate from final evaluation. A rolling-origin procedure can estimate weights from forecasts whose outcomes were observed before \(t\), issue the combination at \(t\), and then score it when the target is observed. An expanding window uses all eligible past origins; a rolling window forgets older origins according to a prespecified length. Both choices can adapt differently when performance changes.

If component models, transformations, forecast horizons, weight constraints, or loss functions were selected after examining a holdout period, that period has become part of model development. Reserve a fresh test period or use a nested evaluation that keeps every selection step inside the training data.

Compare the equal-weight rule, any estimated or shrinkage rule, and the individual forecasts on the same origins and against the same realized target. Report the evaluation loss and the common sample. Otherwise, a method can appear stronger simply because it was scored on easier dates or a different target definition.

The loss should match the target and question. Squared error evaluates distance under a quadratic penalty; absolute error gives a different weighting to large misses. For volatility forecasts scored against noisy realized measures, proxy conditions and the chosen loss matter; the QLIKE and squared-error guide explains those distinctions. Do not pick a loss after seeing which method it favors.

A forecast comparison test can help quantify whether observed loss differences are consistent with sampling variation. The Diebold–Mariano guide covers paired forecast-loss comparisons and the assumptions that matter when forecast errors overlap or remain dependent. A test result does not repair look-ahead bias or prove economic value.

If you tried many component sets, weight constraints, or loss rules and kept the version that looked best on the same evaluation period, the reported comparison also reflects that search. The White Reality Check guide and probability-of-backtest-overfitting guide discuss selection across many trading rules; neither replaces a clean forecast archive or a preplanned evaluation.

Use a validation design that reflects the intended horizon. If daily forecast origins predict overlapping multi-day outcomes, successive losses can be dependent. Account for that dependence when estimating uncertainty, and do not treat a large number of overlapping rows as the same number of independent experiments.

Before combining, check what each model predicts. A conditional mean, a variance, a quantile, and a full probability distribution are different forecast objects. A linear average of point forecasts does not automatically produce a calibrated density forecast or a reliable tail-risk estimate.

For strictly nonnegative quantities such as variance, inspect combined values before applying them. A convex average of nonnegative forecasts remains nonnegative; a fitted regression with negative weights or an intercept may not. Clipping negative values at zero changes the forecast rule and should be specified before evaluation.

The number of forecasts matters too. Adding weak or nearly duplicate inputs can increase the number of fitted weights without adding much new information. If the historical archive is short relative to \(K\), prefer a transparent low-dimensional rule and show that the conclusion is not driven by a single component.

Combining model forecasts is separate from combining trading positions. A lower statistical loss does not show that a strategy using the forecast will earn more after fees, spread, slippage, market impact, and risk limits. Forecast construction and the trading rule that consumes it need separate, out-of-sample evidence.

A useful report names the forecast target, horizon, units, component models, origin timestamps, archive period, and missing-data rule. It states whether weights are equal, estimated, constrained, shrunk, or allowed to vary over time, and how every tuning choice was made.

Show results for the equal-weight benchmark, the selected combination, and the strongest individual forecast on the same evaluation dates. Include the loss function, uncertainty method, and any treatment of overlapping outcomes. If weights move sharply between windows, show that instability instead of presenting only their final values.

Finally, separate the empirical claim from the decision claim. “This combination had lower average loss in this sample under this loss function” is narrower and more reproducible than “combining forecasts improves trading.” The first can be tested from a saved forecast archive; the second requires a strategy, execution assumptions, and separate evidence.