The Model, in the open
Most "deal scores" on the internet are a vibe with a number painted on. This one publishes its workings. Bump Score is a learning model with three data diets and a feedback loop — and this page reads its live state, not a brochure copy of it.
VERSION —
PREDICTIONS LOGGED —
FARE OBSERVATIONS —
OUTCOMES REPORTED —
What it trains on
| Signal | Source | Status |
|---|---|---|
| Airline denied-boarding history | U.S. DOT Air Travel Consumer Report, Jan–Sep 2025 (voluntary + involuntary, per 10,000 passengers) | ✅ Fitted in v1.0 |
| Schedule structure | Published schedules in live fare data: last flight of day, hub banks, evening banks, peak dates, trunk routes | ✅ Structured priors in v1.0 |
| Fare movement | This site's own trackers re-price every loaded flight daily (Duffel observations stored in the model database) | ✅ Collecting — 2 pulls on day one; signal weight grows as history deepens |
| Ground truth: did it actually oversell, and what was paid? | Traveler outcome reports from the "Report the outcome" button on every flight card | ✅ Pipeline live — retraining blends outcomes in once 20+ reports exist |
Current weights — served live from the model store
Loading the live scoring config…
The loop
- Predict. Every position the terminal renders is logged to the model database with its score, factors, fare and model version. No silent edits after the fact.
- Observe. Daily fare trackers re-price the book, so the model sees sell-through pressure build (or not) before departure.
- Verify. Travelers report what actually happened at the gate — oversold or not, volunteered or not, what was paid.
- Retrain. A retraining pass blends reported airline oversell rates with the DOT prior (a pseudo-count of 20 keeps small samples honest) and writes the updated scoring config back to the store the site reads. The pass runs on schedule and its result is recorded in the basis note above.
What the model is not
Airlines do not publish real-time oversell or load data, and this model never pretends to have it. It estimates from published history, schedule structure, fare behavior and reported outcomes. Early on, the DOT prior does most of the work and the counters above will say so. As outcomes accumulate, the model's center of gravity shifts from published history to verified ground truth — that shift is the product.
Advertisement slot — affiliate travel offers (disclosed; never affects the model)