Track record
Every forecast Gold Desk shows is logged and scored against what the market actually did. Good and bad results are both shown, as they come in.
How to read this
68% / 95% held: how often the real price ended inside the range. A well-calibrated forecast lands near 68% and 95%. Much lower means the ranges were too narrow; much higher means they were too wide.
Interval score (Winkler), in % of the price: the range's width plus a penalty for every miss, proportional to how far outside the price landed. Lower is better. A range can't look good here by being too wide or too narrow.
Error vs no change: average distance between the expected price and the outcome, next to the same distance for "the price stays where it was". A model with no edge shows the two nearly equal, and ours mostly do.
Homepage & chat: next 15 minutes to 2 hours
Loading…
Gold
Silver
Chat estimates: 1 day and longer, Karachi 24K
Each estimate is checked against the Karachi 24K rate on its target day, so this includes what the model can't see: rupee moves and the local premium.
Offline test (before going live)
Scores on the untouched holdout from the last retrain. Status ok means the model beat "no change" with statistical significance; no edge means it didn't, so no direction is shown for that horizon.
Method: how the forecasts are built.