Several models weighted by how well each has forecast so far.
The alternative, picking the one model with the lowest past loss, was
measured on 15 series: it beat the default on 8 and cost up to +47 % QLIKE
where the ranking flipped (VNDA, SMMT). Weighting every model by 1 / its
mean past loss averaged −5 % with a worst case of +4 %. At each origin only
windows that have already ended count, so no weight has seen the future.
Several models weighted by how well each has forecast so far.
The alternative, picking the one model with the lowest past loss, was measured on 15 series: it beat the default on 8 and cost up to +47 % QLIKE where the ranking flipped (VNDA, SMMT). Weighting every model by 1 / its mean past loss averaged −5 % with a worst case of +4 %. At each origin only windows that have already ended count, so no weight has seen the future.