The old card showed one cryptic “79% accuracy” badge. It actually grades the 8am morning baseline ($2,478), not the live $1,980 number sitting right next to it. Below: the current card for reference, then two ways to fix it. Numbers are illustrative mock data in the same shape the API already returns.
↑ The “79% accuracy” gives no hint of what it measures, over what window, or that it’s grading a different number than the one shown.
↑ All three numbers always visible, each with a one-line plain-English caption. Taller card, zero clicks.
↑ Card stays compact like today; one tap reveals the full breakdown. Tap the “accuracy ▾” chip above to try it.
Yesterday Did the model nail yesterday? The single most recent predicted-vs-actual. Your “daily accuracy.”
Morning forecast On an average day over the last 30 days, how close was the 8am baseline to where the day finished? This is today’s “79%.” It’s blind to how the day actually unfolds.
Live projection On an average day over the last 30 days, how close was the continuously-revised number (the big $1,980) to the final total? This is the accuracy of what you’re actually looking at — and it’s normally higher, because it learns from the day as it goes.
All three are 100 − average error. Freak days (way above/below) are excluded so one weird day can’t swing it. A run of soft days like this Friday is what pulls the morning number down.