A service that verifies itself

All always-on verification tracks (38)

The six above are representative examples — in fact every track below is recorded at publication and scored automatically after its set period. We publish the scoring criteria first to keep ourselves from changing the criteria after the results are in.

A single track is scored at multiple horizons, each on its own — being right after one day and wrong on the weekly horizon really happens, and a record must keep that too. Tracks that look at only one horizon list just that one. When horizons are added, this table changes first.

Data methodology, limits and corrections policy

This section describes how records are made and graded, the limits of that process, and how errors are corrected. Formulas, weights and threshold values are not published.

Record → grade → aggregate

1. At publication we record the symbol, direction, price and the basis for the call.
2. Each item is graded against the actual market result at its own set points (1 day, 3 days, weekly or monthly, depending on the item).
3. If a point is not graded within the grace period after it falls due (1.5 days for points shorter than 30 days, 3 days for points of 30 days or more, such as monthly), it is left as missing and never back-filled later.
4. For items that grade direction, the shared flat band is 1% from the publication price. Moves inside it are classed as flat and excluded from the hit-rate denominator.
5. Graded results are collected, with sample counts, in the "Full grading results" table above.

Limits

• Where there are fewer than 20 samples, no "criterion met" or "criterion missed" label is attached even if a rate is shown.
• Past grading results do not guarantee future results, and they are not returns.
• Grading covers only what a publication itself claims. Trade levels are judged on direction and reaching the target, watch prices on whether the first target or the invalidation price came first, and inflection points on whether the level held. We do not grade things a publication never said.
• When points are missing, the sample count can differ between points of the same item.
• Individual records for content-judgement items are cleaned up after 90 days.
• The "record anchoring" above only shows that the list existed at that time; it does not mean the records are correct.

Corrections policy

• When we find and fix an error in grading or aggregation, we record it with the date under "Bug fix" in the update log. View the update log
• There is not yet a dedicated screen that collects only corrections to grading results, so no correction count is shown.

Statistics disclosure policy

Scoring results (such as directional hit rates) are shown as they are in the “All scoring results” table above, with their sample counts and including weak ranges, even when samples are few. Early rates from few samples are easily swayed by chance, so read them together with the “too few samples” mark. Performance figures such as the Sharpe ratio, win rate and average R are not shown until there are enough samples and statistical significance.