4W CBB V1.9.18 MIDNIGHT BLUE • BLACK • WHITE
V1.9.18 Projection Formula Repair Test Build: the game-day workflow stays visible while advanced rating tools, diagnostics and historical research are tucked into expandable sections. Includes an embedded preseason rating package dated 2026-11-01; its source, team-division membership and roster methodology are NOT independently verified by this file. It must not be described as verified roster-aware coverage. Historical 2025–26 Torvik proxy adjustments are retained as research references but are OFF by default. Betting recommendations remain hard-locked pending a verified D-I roster feed and 4W-specific chronological held-out validation. Missing sportsbook totals are never treated as zero; genuine model-vs-market total disagreements are flagged. Projections are never forced toward sportsbook lines. The CBB Odds API key belongs only in the dedicated Cloudflare Worker secret named ODDS_API_KEY; never place it in this public HTML. Empirical calibration is not certified.
1. Load Division I schedule
2. Team ratings & automatic projection inputs
Load the 2025–26 Bart Torvik baseline from your dedicated CBB ratings Worker. Verify the Worker response and source before treating it as validated input. The 2026–27 blend and recent-form adjustments will be enabled only after current-season game data is available and validated. AdjO/AdjD are points per 100 possessions; tempo is possessions per game.
Advanced Rating Tools Optional CSV override & roster diagnostics
Built-in 2026–27 roster ratings — optional CSV override
An embedded 2026–27 preseason ratings package is loaded automatically (dated 2026-11-01). Its source provenance and complete Division I membership are not independently verified, so treat it as provisional preseason test input—not a verified roster-aware feed. The 2025–26 Torvik ratings remain available as the prior-season reference. CSV import below can override matched teams when you have a dated source with compatible AdjO, AdjD and tempo. Independent SRS belongs in Advanced Audit and never changes projections. No player values or transfer/freshman bonuses are invented.
CSV columns: team, adjO, adjD, tempo, returning_minutes_pct, returning_production_pct, transfer_in_count, transfer_out_count, freshman_count, coach_change, source, as_of. AdjO/AdjD/tempo must be the provider's roster-aware team projections; continuity fields are retained for audit, not converted into arbitrary point bonuses.
Advanced audit — Independent preseason SRS comparison
Independent preseason SRS comparison (does not change projections)
Compare an independent roster-adjusted preseason SRS source against the Torvik baseline. SRS is a separate rating scale, so it is stored and reported separately—not converted into AdjO, AdjD or tempo, and never used to change the Monte Carlo inputs.
CSV columns: team,srs,srs_off,srs_def,source,as_of. The Lab Analytics preseason ratings page describes its SRS as adjusted for returning production, incoming transfers and recruiting class. These values are not interchangeable with Torvik AdjEM/AdjO/AdjD. Import only a dated source export with team names and the full D-I team universe; partial top-50 extracts are rejected.
Away team
Home team
V1.9.16 Projection Formula Repair & Calibration Safety Controls — TEST ONLY
Research reference only — not active by default: the prior exploratory 2025–26 Torvik proxy fit reported total -14.3 points and home-margin -3.9 points. Those values are not verified against archived 4W forecasts, are not certified for the current preseason ratings package, and must not be called calibrated 4W coefficients. The controls below start at zero. Use nonzero values only for sensitivity experiments, never as validated corrections. The recommendation gate remains locked.
Calculate a projection to run automatic safeguards. These are conservative review flags, not empirically calibrated betting thresholds.
Projection Audit Diagnostic inputs & formula
Projection audit — diagnostic only (V1.9.18)
Shows the exact rating inputs and formula components. It does not certify forecast accuracy.
3. 4W Matchup Analyzer — test build
Side-by-side comparison reads the exact same live team-rating map used by the projection engine and refreshes automatically after ratings load or projections calculate. National ranks below are calculated from the teams present in the loaded ratings map; they are not a separate official ranking feed. Lower AdjD is better. This first CBB version compares adjusted offense, adjusted defense and tempo only—shooting splits, turnover rate, rebounding rate and free-throw rate are not included because the current rating feed does not supply those fields.
| Metric | Away team | National rank | Home team | National rank | Edge |
|---|---|---|---|---|---|
| No matchup data loaded. | |||||
Important: the comparison is descriptive, not a pick. It does not modify AdjO/AdjD, projected scores, Monte Carlo probabilities, expected value or recommendation eligibility. We will only connect matchup adjustments to the projection engine after independent validation shows they add predictive value without double-counting.
4. Sportsbook lines
5. Monte Carlo simulation
Monte Carlo Cross-Check Sampling math diagnostic
Monte Carlo cross-check — analytical approximation
Compares simulation frequencies with the theoretical normal-model probabilities implied by the same means, standard deviation and correlation. A close match checks sampling math only; it does not validate the assumptions or the team projections.
Market value estimates
Market signal review — experimental
Reviews model-versus-market calculations; it does not issue picks. Mathematical candidates remain BLOCKED until current-season roster coverage and held-out calibration are verified. A PASS is expected during this stage. Default screen: at least 3% estimated ROI, 2.0 percentage points above break-even, and each market side must independently be -160 or better. One excluded price does not block other markets.
Historical Score Baseline Audit Descriptive only; not calibration
5A. Historical score baseline audit — descriptive, not model calibration
Import a completed-game CSV with game_date, away_team, home_team, away_final, home_final. This reports actual score distributions only. It does not claim to validate projections because genuine pregame forecasts and pregame rating snapshots are not contained in a final-results file.
Coverage warning: the supplied SportsDataverse schedule export may include non-Division-I opponents. This tool cannot certify Division I membership from the five required score columns. Do not interpret these descriptive averages as forecast error, model calibration, or betting performance.
Historical Backtesting & Accuracy Validation Holdout testing & data-quality gates
6. Historical backtesting & accuracy validation
V1.6 validation gate: no model version is labeled validated until it is tested on held-out historical games using only information available before tipoff. Store predictions, final scores, closing lines, venue, roster snapshot date and data-source timestamps.
Historical Data Import Workbench Combine three verified files
Historical data import workbench — combine three verified files
Use this when projections, final scores and sportsbook odds come from separate archives. Upload one file for each source. Rows are joined by exact game date + normalized away/home team names. source_game_id is retained in the source files for audit, but is not used to guess a match. Ambiguous or duplicate joins are rejected, never fuzzy-matched.
V1.7.3 template separates model projections from sportsbook lines: pregame_home_margin_projection and pregame_total_projection are model outputs; pregame_home_spread_line and pregame_total_line are market lines captured before tipoff; closing_home_spread and closing_total are verified closing lines. Required timestamps must prove inputs were captured before tipoff. The last 30% of games by date is held out and scored separately. ROI/CLV are not inferred without timestamped odds and a locked bet record.
Forecast Tracking & Export Local history and CSV exports
7. Forecast tracking & export
Predictions are stored locally in this browser. Export regularly. A historical backtest must use pregame-only inputs and archived odds; do not confuse projection accuracy with betting ROI.
| Date | Matchup | Model margin | Model total | ML | Cover | Over |
|---|
Build Notes & Validation Gates Model limitations & release notes
8. Build notes & validation gates
- V1.4 is the interface and calculation foundation, not a validated betting model.
- The dedicated CBB Worker retrieves ESPN schedule events; schedule data is not a verified efficiency dataset.
- Team rating CSV must be from a consistent, documented source and include all Division I teams.
- Roster turnover, injuries, transfers, freshmen, coach changes, rest, travel and neutral courts must be modeled before live use. Baseline-only forecasts are explicitly tagged in the validation log.
- Confidence grades A/B remain locked until chronological held-out validation is calculated on at least 200 completed holdout games with probability calibration and data-quality checks. Simulated edge alone cannot earn a high grade.
- Backtest at least 200–500 games using historical pregame data. Historical ROI requires archived closing lines and results.
- Never deploy this Worker or page over the NBA app. Use a dedicated CBB Cloudflare Pages project and CBB Worker.