
Post by
DonJuan21 | 2019-08-02 | 06:40:32
It was interesting getting the exponential decay constant right. If it is too gradual, it shows oscillations when the forecast "n" get's slim, like about a 48hrs out.
To make the blend from forecast to forecast not be oscillatory and model realistic non-oscillatory transitions, it can't take much more than 24hrs into account.
That is, the half-life reliability of any particular forecast is about 4-5 hrs. A 24hr old forecast can't be trusted at the 3% level. A 6hr old forecast is worth about 40% confidence.
Wish we could remove those discontinuities.
You can see that when you play the game. About half the time, there's a noticeable shift as forecasts transition.
DJ
More Cow Bell, more resolution.