South Ark Daily Picks / Model Methodology

Model Methodology

Original baselines, chronological evaluation and explicit abstention rules.

Professional score model

Separate NBA SADP-NBA-context-ridge-2.0 and NHL SADP-NHL-goal-count-2.0 models fit scoring separately from market odds using ridge regression with fixed penalty 20. Each example includes own recent scoring, opponent recent points allowed, venue and rest inputs where recorded. The last 12 games shrink toward the league scoring baseline with two pseudo-games. NFL, NBA, NHL and MLB fit separate coefficients and residual distributions; college football retains its separate SE v3 framework.

All games on the same calendar date are predicted before that date’s outcomes enter training. Held-out margin residuals from earlier games estimate win probabilities after a minimum sample of 40. Last-season rolling-origin validation reports score, margin and total error, Brier score, log loss and calibration. Current forecasts require verified future kickoff, at least five recorded games per team and completed history within 180 days.

NBA and NHL add short-term form, venue scoring splits, verified rest, back-to-back indicators and prior-opponent strength, computed exclusively from earlier results. NHL uses independent Poisson goal counts for outcome probabilities, with tied margins split equally for overtime; overtime/goalie effects are not separately fitted. NBA probabilities use earlier chronological margin residuals. These are result-based research models, not full pace, lineup or goalie models. Missing advanced inputs stay unavailable. NBA and NHL fit private verified league results separately when available. Schedule coverage checks all 30 NBA and 32 NHL teams; daily outputs are saved once per Central calendar day. Private results are refreshed by bounded owner stages; a failed source retains the earlier snapshot. NBA player statistics remain private research inputs; NHL player/goalie logs remain unavailable. Team box-score summaries are descriptive and do not introduce unvalidated lineup adjustments. Expected goals, possession efficiency, travel distance, starting goalies and injury adjustments remain unavailable. MLB pitcher-aware run-distribution modeling is not enabled.

Player baseline

NFL and NBA categories use recent observed current-season player-game logs when available. The last ten recorded games shrink toward season mean with two pseudo-games; at least four current-season records are required. A threshold estimate is the observed fraction strictly above the user-selected threshold. Anytime touchdown uses recorded rushing and receiving touchdowns, not passing touchdowns. Role, injury and opponent adjustments are unverified. These descriptive estimates have not been backtested for calibrated player betting probabilities.

Immutable history

Every automated professional game forecast is content-hashed and stored as an append-only private object before publication in the ledger. Material changes produce a new revision. Original market odds and lines stay null without verified offers. Final outcomes are separate grading records. Retrospective backtests never appear as forward picks.

Confidence and daily publication

Confidence scores remain unavailable until calibrated against sufficient held-out observations. We do not translate an unvalidated win probability into an Elite Edge label. Each owner odds stage saves a content-hashed daily report; visitors read cached source observations. Daily reports may contain zero qualifying plays. Descriptive model-versus-price differences are percentage points, not proof of profitability. Original priced selections stay locked; later report revisions are new objects.

Selection policy

No professional betting selections are enabled. Price verification, adequate chronological validation, input-quality checks and a predeclared selection threshold are required. A fair probability is not proof of profitability.