Forecast track record

Updated: October 14, 2026 · Author: Jennie Chu · Reviewed by: Gemral Research Desk · Editorial Policy

What does the Gemral Edge forecast track record measure?

Each Gemral Edge forecast is a probability that a pre-defined event happens within a fixed window, for example that a stock returns more than its sector ETF over the next 30 calendar days after a member of Congress discloses a purchase. It is not a price target and not a trade recommendation. The track record scores every forecast after its window closes, including the misses, so readers can judge the model on its full history instead of on selected examples.

Why does the page not show a single accuracy percentage?

A bare hit rate is easy to inflate. If an event only happens one time in ten, a model that always says it will not happen is right ninety percent of the time and still useless. The page therefore reports the base rate, meaning how often the event happened with no model at all, next to every number, and it uses three measures that cannot be gamed that way: calibration by probability bucket, the Brier score, and the Brier Skill Score against the base rate.

How are calibration and skill computed?

• Calibration: forecasts are grouped into probability buckets and the average forecast in each bucket is compared with how often the event actually happened. A well-calibrated model sits on the diagonal.
• Brier score: the average squared gap between the probability and the outcome. Lower is better, but it means little alone, so the Brier of the base rate is always shown beside it.
• Skill score: one minus the model Brier divided by the base-rate Brier. The reference rate for each year uses only earlier years, so the comparison is point-in-time. The score is hidden until at least thirty forecasts are scored.

What is the difference between live forecasts and point-in-time replays?

Live forecasts are published before the outcome is known and scored after the window closes. Point-in-time replays recompute history using only data that was public on each forecast date, which gives a much larger sample, but they are kept in a separate ledger and are never presented as live results. Each scored forecast has a public receipt with the ticker, the forecast date and the outcome.

What do the current results show?

The current production model does not show measurable skill over the base rate on point-in-time replays, and the page says so. Many tickers are still compared with SPY because they have no sector ETF mapping yet. Gemral Edge publishes these results rather than hiding them, because the value of the ledger is that it can be checked. Nothing on the page is investment advice and past results do not guarantee future results.

How is a forecast recorded before the outcome is known?

A producer job runs every day and records each new forecast in the ledger with the ticker, the outcome definition, the probability, the benchmark and the date. The database only accepts a live forecast dated today and only accepts a replay dated in the past, so a live row cannot be back-dated. A separate resolver job closes each forecast once its window has ended and the price data for the window is available, and it records the outcome next to the original probability. Because the forecast and the result are written by different jobs at different times, the ledger cannot be filled in after the fact with numbers that already know the answer.

Why compare against a sector ETF instead of only the market?

A defense contractor that rises with the whole defense sector has not done anything special. Comparing each stock with the ETF of its own sector removes most of that shared move, so a forecast is only counted as right when the stock did better than its peers, not just better than cash. When a ticker has no sector mapping yet, the ledger falls back to SPY and records which benchmark was used, so readers can filter those forecasts out.

How should a reader use these numbers?

• Look at the sample size first. A skill score from a few dozen forecasts moves a lot from one month to the next; the page hides the score until the sample is large enough.
• Read live and replay results separately. Replays show how the method would have behaved on past data; only live rows show what was actually published in advance.
• Check the receipts. Each scored forecast links to a receipt with the inputs and the outcome, so any single number on the page can be traced back to rows.
• Compare with the base rate, not with zero. A model that only matches the base rate adds no information, even when its hit rate looks high.

The same ledger is available to API customers through the forecast, track-record and receipts resources, so a fund can rebuild every chart on this page from raw rows and apply its own tests.

What related alternative intelligence streams are available?

Cross-reference correlated public records across our alternative intelligence directories, statutory trackers, and quantitative research libraries:

• How the numbers are computed
• Developer API and bulk history