Yesterday matters less

A result from three years ago says less about a side than one from last month. The fit gives every past match a weight that decays with its age:

w(t) = e^{−ξ (T − t)}

where T is today, t the match date and ξ the decay rate. It is easier to think in the half-life: the age at which a match counts half. Our backtests tried 120, 180, 240 and 365 days; 240 days priced best against the closing line, and that is what the platform uses.

Three half-lives on the same axis. At 240 days a match from last season still counts about a fifth as much as one from last week; at 120 days it has almost gone.

Why not shorter?

A short half-life makes the model reactive: it chases form and prices last week's 4–0 as a new normal. A long half-life makes it stubborn: it prices last season's champion as a champion in October. The backtest is the referee, and it prefers the middle.

Refitting

Because the weights change every day, the model is refitted often. On the platform the weekly refit is what the walk-forward backtest imitates: fit on everything before the week, price the week, move on.

Below, move the half-life and watch how much of the last three seasons the fit is really looking at.

Why not shorter?

Refitting

3 questions