How our daily picks are produced
We combine independent sports forecasts, available sportsbook prices and sourced research to build a list for human review. The process can return no bets. All three sports remain experimental for betting decisions; a completed review does not establish a profitable edge.
The simple flow: Refresh sports data and prices → run the models → research the slate independently → combine the candidate pools → check the price and sourced context → publish up to 10 reviewed ideas → you decide and track.
When the work runs
These are scheduled starts, not guaranteed publication times. Jobs can queue, feeds can fail and books can post markets later. A research job waits for the relevant sports refresh when triggered by that refresh; it does not assume that a scheduled job has finished. The quote timestamp in the picks table is the sportsbook’s timestamp, never the page refresh time.
| Stage | Scheduled start, Eastern | What happens |
|---|---|---|
| Early morning / editorial scheduler | daily: 5:07 a.m. daily: 6:07 a.m. daily: each hour from 5 a.m. through 9 a.m., at minute 27 and 47 daily: 11:07 a.m., 4:07 p.m., 9:07 p.m. daily: hourly at :17 | Conditional feed recovery, price rundown and separate daily articles. Morning recovery attempts are eligibility-gated; each trigger does not mean a new paid article. |
| NFL data and models | Wednesday: 10:00 a.m. Thursday: 5:00 p.m. Sunday: 8:00 a.m., 11:00 a.m., 8:00 p.m. Sunday: 12:30 p.m. Sunday: 3:55 p.m. | Refresh statistics, player estimates, scoring projections and available prices; publish before research. Also callable by early morning recovery. |
| MLB data and models | daily: each hour from 10 a.m. through 10 p.m., at minute 15 | Refresh history/training cache, prices, lineups and forecasts. Also callable by morning recovery. |
| NHL data and models | daily: 10:30 a.m., 4:30 p.m. | Refresh regular-season statistics and prices, run inference, publish all qualifying candidates. Later refresh can capture newly posted props. |
| Top Picks discovery and review | daily: 8:45 a.m., 10:45 a.m., 12:45 p.m., 3:45 p.m., 4:45 p.m., 6:45 p.m. | Runs after successful sports refreshes as well as these checks. Independent discovery, model screening, sourced assessment and publication share the daily research cap. |
The early editorial scheduler can request fresh NFL and MLB data from 5 a.m. through 9:59 a.m. when its board is missing or at least 75 minutes old. The same morning freshness rule applies to MLB, independently of whether daily-article writing has finished. MLB also refreshes hourly from 10:15 a.m. through 10:15 p.m. An early run can still have no candidates when books have not posted markets or a feed fails. NHL’s first scheduled refresh is 10:30 a.m. Props not posted then can enter at 4:30 p.m. Current market coverage and unavailable feeds remain visible; missing odds are never replaced by guesses.
The order inside a successful run
- Refresh sports data and prices. The sports job retrieves its configured statistics, schedule and authorized odds feeds, retains stable identities and quote times, and archives inputs. This is not a claim that every injury, lineup or weather observation is a numerical model input.
- Run the sport’s model pipeline. Current inputs produce independent quantity forecasts and outcome probabilities where supported. NHL uses versioned trained models; MLB refreshes its cached daily training bundle as needed; NFL refreshes player/history estimates and its scoring projections. Missing models stay missing. The data and model update finishes before the review job consumes its published board.
- Research the slate independently. A separate AI-assisted search sees the available games and market types, without our model predictions or rankings. It looks for opportunity, matchup and participation questions. Suggestions must resolve to a real current offer with the exact game, player, market, side, line and book. Search suggestions are unverified leads until the underlying dated reporting is retrieved. Large slates are processed in bounded batches, with games submitted and source coverage recorded separately. A game being included in a search request does not mean every market or report was researched.
- Build the union of candidates. Keep offers passing the sport’s quantitative screen, plus independently researched offers. Neither route must agree with the other. Model quality gates, quote freshness and exact identities apply. The best available price is selected for an identical outcome and line. There is no four-pick quota and no one-bet-per-game rule.
- Collect evidence and investigate the case. Retrieve dated league/team reporting, injury information and relevant search leads. Check expected opportunity, starters, lineups, opponent effects, weather and settlement where evidence is available. Compare the projected quantity with the offered line, the model probability with the actual price, and the offered threshold with other books. Follow-up questions and unresolved source failures are retained. Professional betting recommendations are opinions, not verified support by themselves.
- Assess and rank. Each completed assessment says consider, wait for a specific material fact, or pass, with its case and countercase. Current consider assessments precede pending/wait/pass rows across all sports, so completed MLB or NHL opportunities remain visible above unreviewed NFL rows. Within an assessment status, each sport retains its own reliability and numerical ranking below; scores are not compared across sports. Related bets are labeled as shared exposure. No joint-bet probabilities are obtained by multiplying marginal chances.
- Build a manageable daily card. Show up to 10 current reviewed ideas, aiming for five to ten when the evidence supports them, with no minimum. Only current consider assessments no older than three hours or explicit analyst selections qualify. Explicit analyst selections take priority. Consolidate alternate thresholds and equivalent bets; keep their original forecasts intact. Wait, pass, unreviewed and changed-price assessments stay in the separate collapsed research pool. Additional reviewed opportunities also remain there. Multiple bets in one game are permitted but carry shared-exposure labels. This card limit is a workflow choice, not evidence of profitability.
- Publish the review record. Preserve the original model, offer, source times, review, exclusions and budget usage. The table shows the actual book’s line and odds, quote time, model and market comparison, review status and tracker action. An independent research candidate without an eligible model says “Not established”; it receives no invented probability or EV.
- Recheck and track. Later sports refreshes can add newly posted markets and changed quotes. Changed model/price inputs or newly collected evidence require reassessment. Unchanged price, forecast and evidence may reuse an assessment for up to three hours, preserving its original timestamp. A changed offer never silently inherits approval. Human bets can be logged in Bet Tracker; the site does not place wagers. Automatic grading is not connected for every supported tracker market.
The formulas and selection rules
Let d be decimal odds, p the unconditional chance of winning and u the chance of a refunded push. Then loss probability is 1 − p − u.
- Expected return per dollar:
EV = p × (d − 1) − (1 − p − u) = p × d + u − 1. - Fair decimal price:
d_fair = (1 − u) / p. A price’s break-even probability, conditional on no push, is1 / d. - Two-way de-vig: for paired opposite prices,
q_i = (1 / d_i) / [(1 / d_1) + (1 / d_2)]. Only matching identities, thresholds, settlement and timestamp windows can be compared. A median line is not the book’s expected outcome.
| Sport / route | Quantitative entry and ranking | Market’s role |
|---|---|---|
| NFL player models | Fitted historical-outcome calibration and at least 3% raw-versus-calibrated estimated value. Withhold matched forecasts outside the calibration input range or with possible partial-appearance workload distortion. Rank using the smaller of raw and calibrated EV when both exist; missing raw diagnostics reduce the reliability tier; estimates above 30% return are put in a lower-priority research tier. This is a sensitivity heuristic, not a confidence interval. | Existing prop consensus may include the offered book and is labeled accordingly. It is contextual, not an independent edge calculation. NFL game quantity forecasts do not establish win probabilities. |
| MLB models | Applicable predictive validation passed; supported line range; at least 3% model EV; at least two paired books; best observed same-line price. More than 30% model EV is held for research. Rank by model EV after review/reliability status. Thresholds remain experimental. | Other-book comparison is separate from the independent forecast. Removing the former additional three-percentage-point gap requirement makes EV the price hurdle. |
| NHL models | Independent forecast, verified settlement, at least 2% model EV and a price meeting adverse-scenario minimums. Rank by the existing worst-scenario fixed-fraction log-growth score. Its stake fraction is a ranking device, not stake advice. | Exact paired other-book consensus excludes the offered book. No new model/market blend is fitted by this workflow; no blend is adopted without held-out support. |
| Independent research | Current available offer matched to a researched hypothesis. No eligible model means no numerical probability, fair price or EV claim. Sourced human judgment is evaluated prospectively. | Line and price shopping are inputs to assessment. A recommendation by another outlet or a discrepancy between books does not by itself establish an edge. |
There is no formula assigning arbitrary weights to “model + consensus + qualitative confidence.” Research evaluates assumptions and price sensitivity; it never manufactures a confidence percentage or changes the original model probability. Market Watch remains the separate price-comparison feature.
Freshness, coverage and cost
NFL quotes expire after 90 minutes. MLB quotes and model checks expire after 90 minutes. NHL Top Picks quotes and forecasts expire after 30 minutes, with model inputs checked within 36 hours. Started games, future-dated quotes and failed model feeds cannot masquerade as fresh model picks. The browser checks expiry every 30 seconds and reloads feeds every five minutes.
All Top Picks discovery and review share a $2.75 per Eastern calendar day allowance. It includes search and review requests, legacy same-day charges and conservative reservations for failures with unknown usage. Morning spending is capped at $2.00, preserving $0.75 for noon-and-later checks. That is one daily allowance, not two budgets; unused morning capacity remains available later. Unspent amounts do not roll over. Existing odds-feed costs and the separate daily-article writer are outside this research allowance. No new paid subscriptions are enabled. A request is skipped if its conservative maximum cannot fit; this can leave money unused and candidates awaiting review.
Research batches rotate across sports. Independent discovery ideas receive an early review turn; later runs prioritize previously reviewed offers with changed facts or prices. This schedules research without assigning a higher win probability to those ideas. The main card holds at most 10 reviewed ideas. The separate collapsed research pool initially shows 20 additional offers, with a button to show all; all eligible candidates remain available for research. A research batch limit controls API payload size, not candidate eligibility. Budget and source limits can leave incomplete research. That coverage is recorded explicitly. No system can promise to examine every possible market exhaustively for $2.75, and no daily picks minimum is imposed.
One-day release test: The owner authorized a $10 shared research ceiling for September 27, 2026 only, including spending already recorded that day. It expires automatically at midnight Eastern; the usual $2.75 daily allowance resumes September 28. This does not change the maximum of ten reviewed ideas on the main card.
Ordering the daily card: After the completed-review gate, explicit analyst selections lead, followed by forecast reliability and an expected-log-growth score on a common scale across sports. The calculation uses the NHL ranker’s fixed 0.0025 fraction; it is an ordering convention, not staking advice. NFL uses the lower raw/calibrated estimate where available, MLB uses its win/push distribution and NHL retains its adverse-scenario score. Missing independent models receive no numerical score. Sport names and feed order do not receive priority, and no sport gets a guaranteed slot. This ordering has not been validated as a profitable selection strategy.
What remains to prove
New independent discovery and review decisions are prospective shadow records. We have not shown that this larger candidate pool, new ordering or qualitative intervention improves returns. Grade original forecasts and reviewed selections separately, retain correlated exposure and execution prices, and assess forecast calibration, returns, drawdown and closing-price diagnostics where trustworthy data exist.