DAILY FIVE · MODEL METHODS
How we select and grade bets
November 5 upgrade: context challenger v1
The original ppg-trial-v1 remains a separately identified benchmark. A new shadow challenger adds NBA possession pace and offensive/defensive efficiency; NFL passing/rushing and defensive EPA plus starting-quarterback efficiency; opponent strength; preseason priors; confirmed availability; rest and travel; and football weather. These are research forecasts, not new official picks. No weights, calibration or profitable results are asserted until measured.
NBA scoring starts with projected possessions multiplied by the paired offensive/defensive ratings per 100 possessions. Early-season rates shrink toward documented preseason priors, with shrinkage strength selected on earlier chronological tuning data. NFL performance features and the scoring baseline also shrink toward preseason priors. Separate league regressions learn context adjustments and the relationship to the DraftKings baseline. Unknown availability, stale context or missing provenance causes a hold rather than an invented zero adjustment.
Train on earlier events, tune on later events, then evaluate once on untouched later games. Purge results that were not known before the next partition. Every feature needs a source, retrieval time and data-through time preceding its prediction. Report sample sizes and total/margin MAE and RMSE against both the original scoring model and the contemporaneous DraftKings line. An unavailable historical odds snapshot cannot be replaced with a closing line or a hindsight reconstruction.
Keep challenger win, push and lower-bound probabilities null. Before promotion, a separate chronological calibration and uncertainty study must assess Brier score, log loss, reliability, push frequency and price-specific one-unit ROI with uncertainty intervals against the no-vig DraftKings baseline. Compare variants and account for repeated model searches. Use actual independent holdout references; a small segment record or good training fit does not establish an edge. No automatic promotion is enabled.
Rank shadow discrepancies in units of each league/market’s training residual scale; NFL and NBA raw point gaps are not comparable edges. Only a genuinely validated probability model may rank official value selections by conservative EV, requiring estimated EV ≥3% and conservative EV >0 at the original line and price. Select up to five singles, one per event; publish fewer or zero when evidence is insufficient.
Capture DraftKings near-close snapshots prospectively in the final five pregame minutes where possible. Favorable line movement and same-line price improvement are reported separately; missing captures stay missing. Closing-line movement is a diagnostic, not proof of profit. Original reports, prices and losses remain immutable, and challenger research is excluded from official ROI. View research status and audit →
NBA benchmark documented November 4
The NBA totals-only benchmark uses each team’s scoring environment E=(season points for + points against)/games; projected total=(away E + home E)/2. It does not apply the NFL home-margin adjustment. This is an uncalibrated baseline without pace, roster or opponent corrections. Never label a point discrepancy as EV.
November 1 launch: a prospective trial
The Daily Five segment begins November 1, 2026. No earlier segment bets exist. First results are reported November 2 from verified final scores. Five one-unit singles create five units of paper exposure; original prices and losses remain in the record.
Our opening model, ppg-trial-v1, uses documented season points scored and allowed per game. Away projected points = (away points scored + home points allowed) / 2 − 0.75. Home projected points = (home points scored + away points allowed) / 2 + 0.75. The 1.5-point home margin adjustment is a fixed trial assumption, not a fitted estimate. Totals equal the two projected scores; spreads use the projected home margin.
For each reviewed matchup, compare the model with DraftKings spreads and totals, select the largest positive point discrepancy, then rank the slate by that discrepancy. Publish at most one single per game. Each card records its inputs and sources. November 1 reviews five NFL matchups; it does not claim a full-market search. Points of discrepancy are not expected profit, calibrated confidence or win probability. This simple model is not opponent-adjusted and does not quantify injuries, roster changes, rest or weather. Discuss those risks explicitly.
Accept exact odds displayed on DraftKings Sportsbook or the official DraftKings Network odds board, recording the source and retrieval time. Prices may differ by jurisdiction and change before kickoff. Check the same line and price again before considering a wager. Never backfill after games start. Trial selections do not require a prior results history; prospective outcomes build the record.
The section below describes the separate dk-value-v1 approach, available only when its independent calibration requirements can actually be met.
DraftKings is the price benchmark
Use publicly displayed DraftKings Sportsbook moneylines, spreads and totals. Every pick records its exact line, American odds, game start, direct source and time checked. No paid feed is required. Prices vary and move; a published snapshot is not a live offer. Unverified or inaccessible odds produce a hold, not an invented selection.
Estimate value, then account for uncertainty
Convert odds into implied probabilities. For a complete market, divide each implied probability by their sum to remove the margin proportionally. This is a market estimate, not proof of the true probability. Sport-specific models use opponent-adjusted performance, confirmed availability, starters, rest, travel, venue and relevant weather. Blend with the DraftKings market baseline using weights learned on chronological validation.
Require independent holdout calibration and record the validation reference with every selection. Model probabilities and uncertainty bounds must have evidentiary support. The initial design has not established a profitable historical record. Until that separate model passes validation, do not publish dk-value-v1 selections. This requirement does not block the explicitly labeled ppg-trial-v1 launch.
Expected value per unit risked = P(win) × (decimal odds − 1) − P(loss). Include push probability explicitly. Qualify at estimated EV of at least 3% and conservative EV greater than zero. Rank by conservative EV. These thresholds are initial design choices, not demonstrated optimal settings.
Publish at most five pregame singles, one per event, only before the event begins. Record odds checked within 30 minutes of publication. Include a minimum acceptable price preserving 3% estimated EV at the same line. Correlation across games can remain; no independence or profit guarantee is implied.
Grade the original published bet
Preserve the original daily report and prices. Check final league or team results and the applicable DraftKings rules. Mark win, loss, push, void or pending. A win earns the original odds’ profit on one unit risked, a loss costs one unit, and a push or void returns zero net units. Keep postponements or unverified outcomes pending until settlement is known.
ROI = net units divided by settled units risked, including pushes and excluding voids and pending bets. Report yesterday’s results separately and maintain lifetime totals and a dated archive. Continue checking older pending picks. Correct a conflicting result transparently rather than silently replacing it.
Sources and limits
Sources are linked per pick and result. Review game facts, lineups and distinctive prose before publication. Do not backfill picks after games, manufacture historical performance, change original prices or claim a verified model without evidence. This is a paper tracking report; it does not place bets.