Latest forecast
| Horizon | Target (UTC) | Median | Δ pips | 50 % band | 90 % band | P(up) |
|---|
loading…
Bars at the size the bars picker names — hourly, or the same hourly series aggregated into four-hour or daily bars of the trading week — times in UTC, for the picked instrument, the primary one unless the picker says otherwise. The cone is the picked model's latest forecast, the reference model's unless the picker says otherwise: dashed median, 50 % band (dark) and 90 % band (light); on a coarser chart it is drawn through the horizon that closes each bar. The markers give P(up), the chance that the bar closes above the origin close, at 1 h, 4 h, 12 h, 1 d, 3 d and 1 w — those that fall a whole number of bars ahead, so 4 h and up on H4 and 1 d and up on D1; the horizons table below lists every one of them whatever the picker says. Name a past origin bar in the replay control and the page draws the cone that model issued then over the bars that followed it, and stops updating until “latest”. Before the forward test began the ledger holds no forecast, so the cone at such an origin is reconstructed from the bars up to it: what the model would have said, marked as reconstructed, and no part of the record the verification table scores.
| Horizon | Target (UTC) | Median | Δ pips | 50 % band | 90 % band | P(up) |
|---|
Scored forecast points whose target bar closed in the window, from the score ledger. MAE in pips (0.0001, the picked instrument's); naive is the random walk (last close carried forward); skill is 1 − MAE / naive MAE, positive when the median beats the last close; coverage is the share of targets inside the 50 / 80 / 90 % bands, ideally the nominal value.