Cookies on Macro by Mark.We use essential cookies for sign-in, locale, and security. With your permission, we also use optional cookies for preferences, for analytics, and for advertising measurement with the Meta (Facebook) Pixel. Read our cookie policy.
Combination of several base forecasters - simple averages, weighted blends, or stacking - to reduce single-model risk.
Graph workspace
The canonical ensemble visualization shows individual model forecasts as thin lines, the combined forecast as a thick line, and the realized value as a dashed line. This makes visible the diversification benefit: individual forecasts scatter around the truth, and the combination (sitting near the center of the scatter) is closer to the realized value than most individuals.
Combined MSE
Mean squared error of the combined forecast, computed exactly from the weights and the error covariance set on the page rather than averaged over a sample. Should be lower than most (ideally all) individual model MSEs.
0.4428
squared units
Best individual MSE
MSE of the best individual model. The benchmark that the ensemble aims to beat.
0.6400
squared units
Diversification gain
Fractional MSE reduction from combining versus using the single best model. Positive values confirm the combination is beneficial. It goes NEGATIVE when one model dominates - equal-weighting a strong model with weak ones is worse than picking it, which is the honest result, not an error.
0.3081
signed fraction (negative = worse than the best single model)
Effective number of models
Herfindahl-based measure of weight concentration. Equals M under equal weights; approaches 1 when one model dominates.
5.00
dimensionless
Error correlation: 0.3
How much the models' errors move together; more alike means less gain from combining them.
Increasing makes individual forecast lines move together. The ensemble spread narrows and the combined forecast provides less improvement.
How many forecasts go into the pool; more adds variety but makes the weights harder to estimate.
More thin lines appear. The combined forecast stabilizes (tighter around truth) as M grows, especially when error correlation is low.
Model quality spread: 0.2
How much the models differ in accuracy; zero means every model is equally accurate.
Higher spread makes some thin lines track truth closely while others deviate widely. The optimal-weight combination hugs the accurate models.
Weighting scheme (1 = equal, 2 = invMSE, 3 = optimal): 1
How each forecast's weight is set: equal shares, accuracy-based shares, or the lowest-error mix.
Switching from equal to optimal typically moves the combined line closer to the best individual model. In this toy the weights use the true error covariance, so optimal never does worse than equal; with weights estimated on a short sample and a large M, it can.
How much the variable being forecast bounces around; more movement makes forecasting harder.
Higher increases the amplitude of all lines. The relative improvement of the combination stays similar in percentage terms.