Research methodology v1.0 · November 16, 2026. BBTN compares forecasts with timestamped sportsbook observations. Model gaps, agreement and market movement are research evidence. No calibrated win probabilities or validated expected-value rankings are currently published.
What the board measures
PredictionTracker supplies spread forecasts and dispersion among forecasts. Jacob supplies spread and total projections. The displayed spread blend is an experimental 50% PT / 50% Jacob average. These forecasts may share inputs; statistical independence has not been established. PT dispersion measures model disagreement, not the uncertainty of a game outcome.
Forecast margins are stored as home score minus away score. A positive projected margin favors the home team; the board converts it into the named team’s sportsbook-style handicap for display. Totals are combined points. Full-game markets are kept distinct from first-half markets. A null means unavailable; zero means a genuine zero only when verified. Neutral-site and team-identity checks are required before a forecast is treated as reliable.
Prices, timing and market references
Each sportsbook quote retains its own line, American odds, observation time, source and timestamp precision. The BBTN opening snapshot is our first recorded observation, not a claim about a bookmaker’s first offered price. Quotes are historical and may be unavailable now. Snapshot age does not imply live access or an executable wager.
The featured shortlist requires two conditions: the model still supports the side at the recorded sharp line, and that same sharp book’s point line has moved toward the model since the first saved observation. Circa is the primary reference; Pinnacle is the fallback when comparable Circa observations are unavailable. A better NoVig or bet365 line, or better odds at the same line, is a bonus rather than a requirement. Model agreement and spread size are not additional gates. Cards are ordered by remaining point gap within each market. Extra points at worse odds are labeled with the higher price rather than assumed to improve expected value. The full board retains opening baselines and their model-gap comparisons. This operational screen has not been validated as a profitable strategy. Pushes, exchange fees, commissions, limits and settlement rules must be evaluated separately. No fee-free execution is assumed.
For American odds a > 0, decimal odds D = 1 + a/100; for a < 0, D = 1 + 100/|a|. Raw implied probability q = 1/D. The displayed two-sided no-vig reference is q(side) / [q(side) + q(opposite)], using the same book, observation time, market and threshold. This proportional normalization is a reference method, not proof of a true probability or a model edge. Integer-line references concern decisive outcomes because a push refunds the stake.
Circa and Pinnacle observations are shown separately. The earlier 60/40 reference weighting has not been validated. Movement is only the difference between comparable observations; its direction does not establish who bet, why a book moved, or whether value remains after the move. Historical closing lines are evaluation benchmarks, never inputs into a simulated earlier decision.
Probability and expected value: currently unvalidated
A point gap is not a cover probability. A future probability model must learn from timestamped pregame forecasts paired with outcomes, estimate error distributions, handle integer-score pushes, and validate out of time. A normal distribution or a fixed game standard deviation will not be assumed without evidence. Early-season rates need shrinkage toward appropriate priors, with opponent strength, pace, roster changes and venue considered.
For one unit risked, EV = P(win) × (D − 1) − P(loss). Pushes contribute zero. Relevant fees reduce net payout. On a half-point market, −110 requires 52.38% to break even. That threshold is a property of the price, not a BBTN forecast. Stars in current and archived articles describe editorial conviction; they are not validated probability estimates.
Current diagnostic findings
On the November 16 board, Jacob projects a lower home margin than PT in all 22 paired games; mean Jacob minus PT is −4.48 points. This is a difference between models, not measured outcome error. Check home-court treatment, neutral sites, team mapping, strength and pace adjustments, and spread-size bias. The supplied projection workbook contains output values and constant arithmetic expressions rather than the generating algorithm; reproducibility of the upstream process remains unverified.
A retrospective November 12 check joins 22 Jacob output rows, PT forecasts and the existing tracker’s recorded final scores. Scores were not independently reverified in this audit, and original pregame preservation is not established. These results are descriptive, not a prospective backtest or evidence of profitability.
| Spread forecast · same 22 games | MAE (pts) | RMSE (pts) | Mean actual − forecast |
|---|---|---|---|
| Jacob | 8.75 | 10.52 | +2.96 |
| PT | 8.50 | 10.36 | -1.23 |
| 50/50 blend | 8.27 | 10.19 | +0.87 |
| Market baseline in Jacob source | 8.09 | 9.61 | -1.86 |
For the same 22 games, Jacob total MAE was 12.76 points and mean actual total minus forecast was −3.27 points. One slate is insufficient for calibration, blend optimization, stable error distributions, or ROI uncertainty intervals. Market baselines in the source workbook may differ from later book snapshots. Missing execution odds prevent a defensible ROI comparison for these research selections.
Validation protocol
Preserve every eligible game, including passes and conflicting forecasts, before tipoff. Keep all snapshots and markets for one game together. Use earlier dates to fit weights, subsequent dates to estimate and check probability calibration, and a final untouched later period to evaluate performance. Record exact cutoffs and model versions. Never use revised ratings, future injury information or closing prices in an earlier simulated decision.
Compare PT alone, Jacob alone, the experimental blend, the market available at decision time, and a conservative model–market combination. Report signed bias, MAE and RMSE for forecasts; reliability plots, Brier score and log loss for probabilities; and price-specific returns, drawdowns and closing-line comparisons for selection policies. Compare forecast-error correlations to assess whether a blend adds information. Closing-line value is a diagnostic, not proof of profit.
Freeze selection rules before testing: model-only versus movement-confirmed signals, agreement filters, large-spread screens, top-gap limits and editorial overrides. Account for all attempted variants. Test performance at the price available after confirmation. Do not optimize on the final test period. Use game/slate clustering for uncertainty estimates, and report sample sizes and distinct dates. With too few independent slates, leave uncertainty estimates unavailable rather than presenting unstable precision.
Records and accountability
Research snapshots are stored as structured, append-only WordPress records. Distinct imports create new snapshots; duplicate normalized imports reuse the existing record. The public archive and JSON exports preserve observation time separately from record creation time. Current snapshots originate from manual imports; no automated spreadsheet synchronization or live sportsbook feed is configured. Historical migrations are labeled retrospectively and do not establish pregame availability.
Official selections retain their posted book, market period, threshold, odds and publication time separately from research. Standardized one-unit-risk results describe a paper staking convention unless actual wagers are verified. Shared recommendations by two writers count separately only in a clearly labeled writer ledger; site-level unique recommendations must be deduplicated. Correlated markets in one game are not independent evidence.
The existing five-pick 2026–27 result is too small to establish a repeatable advantage. No probability calibration, blend-weight superiority, validated sharp-confirmation rule or robust ROI confidence interval is currently claimed. Reassess and version this methodology when the generating process and a sufficiently broad preserved history are available.
Research board · Dated archive · Official results · Probability calibration reference · Time-ordered evaluation reference
Reproduce the diagnostic audit
Download the 22-game diagnostic CSV and Python audit tool. Run python audit.py history-audit.csv with Python 3.9 or later. The tool reports bias, MAE, RMSE, paired error correlation and per-date results, and rejects duplicate games. Weight fitting remains unavailable for this one-slate retrospective dataset. The script does not produce calibrated probabilities or ROI confidence intervals.
