How to use the desk

Matches move from Upcoming to In progress at their scheduled kickoff. The original forecast remains available until a result is collected and graded, when it moves to Results.

Results are checked three times daily, so a finished match can remain in In progress with an Awaiting result label. This is not a live score feed. Postponed and suspended matches retain their reported status.

Use the Prediction model filter to switch all Upcoming or In progress cards between Jarvis + history (the default), Original Jarvis and AI learning. Probabilities, signals and selections follow the chosen model; matches without a saved forecast are labelled. Each card names the model you are viewing. Results keep the saved main forecast. Compare models opens the saved Original Jarvis, Jarvis + history and AI views where available.

Home / Draw / Away are probabilities over regular time. Confidence is the highest of those probabilities; it is not a guarantee.

Volatility measures uncertainty across outcomes. Higher values indicate more uncertainty.

Trap, Draw Watch and Upset flag conditions in the model. They can overlap. Neutral means none of those flags fired.

The suggested bet type comes from the model’s probability and risk rules. The priced selection also checks the saved odds and can favour a different outcome.

Watch only keeps the forecast visible but withholds a priced selection. This can happen with missing prices, limited history or unmet selection criteria.

Paper returns grade the saved priced selection, using one hypothetical unit. No bets are placed.

Jarvis + history is the main forecast for new matches across all covered domestic and European competitions. It uses up to 550 days of results, with recent games carrying more weight. Original Jarvis remains a comparison benchmark. Matches already started keep their saved model and probabilities.

The AI learning model already publishes home/draw/away predictions under each match’s Compare models link. Domestic and European competitions have separate trained models and performance records. New forecasts update with the scheduled data refresh; the AI retrains at least weekly. It remains watch only while live results accumulate, with no fixed date or automatic switch to make it the main model. Historical AI results were mixed.

Verified injury and lineup news are not included. Goal-market forecasts come from Jarvis.

Older results provide an early-season baseline, but age and league differences reduce their weight. A team can have many old games and still have a small relevant sample.

Promoted teams receive additional caution. Missing, limited or stale history and league transitions remain flagged; they are not treated as sufficient evidence for a priced selection.

Forecasts freeze before kickoff. Results use the regular-time score, excluding extra time and penalties. Voided matches are excluded from probability and return totals.

Accuracy counts the leading outcome correctly predicted. Brier score measures probability error on a scale from 0 to 2; lower is better. Paper return uses the price saved with each qualifying selection.

The AI lab compares only matches with saved forecasts and settled results from all three models. Historical tests are separate from the live record and do not establish future betting profitability.

Data comes from football-data.org. Predictions publish automatically three times daily; the process starts at 04:17, 12:17 and 20:17 UTC. Calculation and publication follow, and schedules can be delayed.

Domestic fixtures cover the next three days; European fixtures can appear up to a week ahead.

Times on the site use UK time. The header shows the latest publication, and updates more than 18 hours old are flagged.

Prices are saved snapshots, not live quotes. Observed rest can miss cup and international matches outside the loaded competitions. Injury and lineup news are not included.