Every forecast Palanor shows is logged when it is made and scored when its date arrives. The results sit on each signal page, in the Track record panel under the chart, and across all signals at palanor.com/foresight/accuracy.
What gets logged
Once a week, every public signal with at least 20 observations and a recent print has its forecasts recorded: Continuation and Palanor Foresight, exactly as the chart draws them, plus the pattern read Foresight uses inside its blend, with the band and the engine version. A logged forecast is never edited.
Checkpoints
Each forecast is checked at two points, set by how often the series publishes:
- Daily, weekly and monthly signals. 7 days and 30 days.
- Quarterly signals. 90 days and 1 year.
- Annual signals. 1 year and 3 years.
How a forecast is scored
- The outcome. The observation nearest the target date, within half the series' spacing (at least three days). Nothing is interpolated.
- Voids. When no observation lands in that window, the forecast is voided with the reason and never scored.
- Beat no-change. The share of forecasts closer to the outcome than the last value was. Carrying the last value forward is the simplest forecast there is; a mode that can't beat it more than half the time isn't earning its place.
- In band. The share of outcomes inside the ±1σ band. An honest band holds about 68%. Much more means the band is too wide; much less means it is too sure.
- Scaled error. Errors are divided by the series' typical step, so signals of different sizes can be compared.
Hold-out tests
Live results take time to accumulate. Until they do, the panel also shows a hold-out test: forecasts rebuilt from the signal's history cut at past dates, each series holding back its usual publication delay. They use today's revised data and today's lattice, so they read a little kinder than live results. They are labelled "Hold-out test (today's data)" and never mixed with live results.
When the panel appears
The panel appears once five forecasts have been scored. Before then, if live forecasts are waiting, it says when scoring begins. Market prices are scored like any other signal, but only the shares are shown.
The record tunes the engine
The record is read back, not only published.
- Every day, a series on which Foresight has not beaten no change (at least 20 scored forecasts over 26 weeks, under 40% closer than the last value, median error at least 10% worse) is held at its last value for 13 weeks, then tested again. A held forecast says so in its label and in the chart's note. A series whose lattice reads lose while its own-history reads win is read from its own history only.
- Every week, the weights that blend trend, pattern and lattice are re-fitted per cadence on the older half of the record and judged on the newer half. A change is kept only when it cuts the median error by at least 2% on forecasts it never saw, does not raise the mean error by more than 1%, does not lower the share beating no change, also wins on the older half, and at least 28 days have passed since the last change. Switching the lattice on or off for a cadence needs a 5% gain.
- Applied changes appear on palanor.com/foresight/accuracy under Engine changes, dated. Forecasts made before a change keep the weights they were made with.