Research
Does the quoted YES price actually track the real probability of winning? For each city-day, the single snapshot closest to the anchor time below (local trading time) is binned by price; within each band, the actual observed win rate is compared to what perfect calibration (win rate = price) would predict. Built from every Polymarket snapshot captured since 2026-05-22, through 2026-11-06.
Testing/optimizing the SYS forecast blend (the reliability-weighted 4-source average behind the Sys Forecast Agreement line below) needs a city-day with all 4 raw inputs present AND a settled outcome - currently 817 of those (2026-05-29 to 2026-11-05), growing at ~5.11/day. Backfilling doesn't help - SYS itself is already computed for 91% of city-days that have any input at all, so the ceiling is raw source-data availability, not missing computation. At the current rate, that sample would double by 2027-04-14 - a reasonable "worth a second look" date for a full 4-source optimization pass, not a hard deadline (a narrower, single-parameter test of just the MKT cap or the bias blend is usable today on a larger partial-source pool).
X-axis is bucketed into fixed 5-percentage-point bands (e.g. the "48%" bar covers 45%-50%, not exactly 48%) - each bar's label is that band's rounded midpoint. Mkt Forecast and Sys Forecast restrict the same bins to only the bucket observations where that raw forecast source (no bias correction) also currently identifies that exact bucket as its pick - same bucket_matches_target resolution Market Knows' own MKT/SYS Strategy conditions use live. Tests whether forecast agreement adds real signal beyond what the market price alone already implies, at the same price level. Sample sizes are much thinner than the main curve (only a fraction of bucket-price observations have a matching forecast pick), so read these two as suggestive rather than confirmed wherever a point looks lighter.
When two or more buckets in the same city-day snapshot are priced identically, only one can actually win - so a tied price can't be cleanly credited to a single price level. This excludes every bucket involved in a tie (both/all of them, not just one) and compares the resulting curve to the original. Low-price bands are the most tie-prone (many clearly-losing buckets often bottom out at the same price floor), which is why this mostly moves the left side of the chart.
For each band: if the band's actual observed win rate above holds up, what's the expected profit from buying 1 YES unit at that band's price and holding to resolution, after the entry taker fee? Same bands, same filters, same data as the calibration chart above - recomputed every time it refreshes. Positive means the mispricing more than covers the fee; negative means it doesn't.
Research
When during the day is it actually worth buying? For each city, every bucket priced at 35% or higher is treated as a hypothetical 1-unit YES buy, bucketed by price band (columns, 5-point bands from 35% up) and by hour of day (rows, 7am-11pm on the event day itself - 11pm is about an hour before that day's close, not the night before - nearest snapshot within 30 minutes of the hour). Cell color is the total realized PnL, net of the entry fee, from making every one of those bets across the full history above.
Updates live: unlike the calibration charts above, this recomputes directly from the raw snapshot history on every page load - there's no cached daily snapshot, so it's always current with whatever's landed in the database so far, at the cost of taking a few seconds to load. Hours are each city's own real local time (e.g. KATL's rows are Eastern, KLAX's are Pacific) via each market's actual timezone - not this server's clock and not a single shared reference time, so an hour label means the same thing everywhere: what the clock on the wall in that city actually read.
Does the 50%-70% price band's edge actually cluster around a specific time of day in each city, or does it just look that way from a handful of lucky or unlucky days? We're checking by pooling every bet in that price range at each hour, giving each hour far more observations than any single heatmap cell above. A real time-of-day effect should show up as a hill spanning several neighboring hours; an isolated spike with quiet neighbors on both sides is the sign we're looking at noise, not signal.