Every forecast site says "trust the model." Here is exactly what ours has and hasn't proven: 77.6% winner-vs-dud accuracy across 58 verified past drops, against a 52% coin-flip baseline — and a live cohort that currently reads 36 forecasts tracked, zero resolved. Both numbers are on the scoreboard. This piece is the plain-English audit of what they mean.
The discriminator behind the Pre-Drop Forecaster was fit on drops that had already resolved — products where the resale outcome (did it clear 2x retail, or didn't it) was verifiable from sold-comp history. We then asked one binary question per drop: knowing only what was publicly visible before release, does the model call winner or dud correctly? It got 45 of 58 right. The 52% baseline matters: always-guessing-"winner" on our tracked set is nearly a coin flip, so 77.6% is signal, not survivorship dressing.
The calls we're proudest of aren't the winners — they're the saves: high-hype drops the model told you to PASS while the crowd was loudest. A hyped dud costs a reseller twice: retail out the door, then fees and shipping to recover half of it back. The model's edge shows up most clearly there, because hype metrics alone would have said COP — it's the scarcity leg (allocation, print-run tells, restock behavior) that flags the trap. When we publish a PASS on something loud, that call is carrying the model's whole reputation.
The live cohort — 36 forecasts published with dated drops, 4 pending resolution windows — hasn't graded yet because grading requires the drop to happen and a clean post-drop comp trail to settle (we wait for the rip-week noise to clear before scoring). A backtest is evidence; it is not proof against the future. The moment live calls resolve, the hit rate goes on the scoreboard whether it's flattering or not, next to the exact forecast we published. That page has no delete button — retired forecasts stay live on this site, permanently, as archived final calls.
Four things this model does not do. It doesn't price black-swan supply moves — a surprise restock or reprint kills a scarcity thesis and we eat the miss. It reads public pre-drop signal only, so channels invisible to it (fashion retail, private allocations) can make a demand score under-read; we saw exactly that pattern on the CDG × AJ11 forecast. Confidence bands are category priors until an item has its own price history — right now most of our sell-window calls run low-confidence for exactly that reason, and the pages say so. And 58 resolved drops is a real sample but not a large one; 77.6% carries an error bar, and we'd rather you know that than not.
The methodology in full detail — data sources, the demand and scarcity decompositions, what the confidence tiers require — lives on the About page. The rolling record lives at /scoreboard. Check back and hold us to it.
The Journal is DropForecast's editorial layer: desk-bylined analysis written on top of the site's own tracked data and models, edited and published by the operator. Desk names are house pen names, not individual staff. Every number cites our own boards, and every call referenced here gets graded in public on the scoreboard.
