Guide: how to actually use this app

One job, five destinations: turn a probability estimate into a safe, honest answer about whether to bet and how much. Nothing here predicts winners — you (or the model, or the automated This Week tab) supply the probability, Research helps you stress-test it and dig into any team, player, or trend yourself, and Tracker is about sizing and survival.

Start here: your weekly checklist (no experience needed)

This section assumes you know nothing about betting math, probability, or any of the terms elsewhere on this page. Just follow the numbered steps in order, every week.

Watch out: This app never places bets and never touches real money. You place every bet yourself, on your own sportsbook, with your own money. Only bet what you can genuinely afford to lose. Every recommendation here is a probability, not a promise — you will lose some bets even when the app is working exactly as intended, and that's normal, not a sign something's broken.

Part 1 — One-time setup

1Click Tracker at the top of the page — it opens on the bet log ("Bets & CLV") tab by default.
2At the top, find the box labeled Starting bankroll. Type in the total amount of money you're setting aside for betting — for example, 2000 for $2,000. This is just a number you type once; the app doesn't connect to any real bank account or sportsbook.
3That's it for setup. This number is saved automatically in your own browser, so you don't need to enter it again.

Part 2 — Whenever you want to find a bet this week

1Click Research, then the This Week sub-tab, then pick a sport (college football or NFL).
2Wait a few seconds. You'll see a message saying it's fetching this week's games — that's normal, it's checking real, live game data.
3Look at the section titled something like "N edges found." Each box below it is one suggested bet, showing a team name, a green BET label, and some numbers.
4You don't need to understand every number. The two that matter most: "Expected value" (bigger is better — it's the app's estimate of how favorable this bet is) and "Recommended stake" (how much money the app suggests betting, based on the bankroll number you typed in Part 1).
5Ignore the collapsed sections further down ("games flagged," "no edge," "skipped") — those are the app specifically telling you not to bet on those games. You can click on them out of curiosity, but there's nothing to act on there.
6Pick one (or none) of the suggested bets. Nothing forces you to bet on every suggestion, or to bet at all.

Part 3 — After you decide to bet

1Go place the actual bet yourself, for real money, on whatever sportsbook or app you use.
2Come back here and click the blue Log this bet in the Tracker button on that same box.
3You'll land on Tracker's Bets & CLV tab with a form already mostly filled in for you. Double-check the "Stake ($)" box matches the amount you actually bet — if you bet a different amount than what was suggested, change that number to match reality. Being honest here is what makes the rest of this useful.
4Click the blue Log bet button. Your bet now shows up at the bottom of the page, labeled pending.

Part 4 — After the game finishes

1Come back to Tracker's Bets & CLV tab.
2Find your bet in the list — it still says "pending," or has already auto-settled if it was logged from a This Week/Research card (see below).
3If it's still pending, click Win if you won the bet, or Loss if you didn't. That's the only thing you need to do — everything else updates on its own.

Part 5 — Checking how you're actually doing (do this anytime)

1On Tracker's Bets & CLV tab, look at the row of numbers near the top.
2Current bankroll is how much money you have right now, based on every bet you've logged and settled.
3Total P&L means total profit and loss — a plus number means you're up overall, a minus number means you're down.
4Win rate is simply the percentage of your settled bets you've won.
5Now look at the chart titled Realized vs. theoretical bankroll. The solid blue line is what actually happened to your money, bet by bet. The dashed orange line shows what would have happened if every single bet had been sized exactly the way the app recommended. If the blue line is running below the dashed line, you've generally been betting bigger amounts than recommended (which raises risk); if it's running above, you've been more cautious than recommended. Neither is "wrong" — it's just an honest picture of what you're actually doing versus what the app suggested.

Tip: That's the whole loop: check This Week → place a real bet → log it → settle it once it's over → glance at your numbers. Repeat every week. Everything else on this page explains the reasoning behind the numbers, and every other tool available, for whenever you're curious — you never have to read it to use the app.

The big picture

The tools are meant to be used together, roughly in this order, whenever you're considering a real bet:

Get a probability. Either your own honest estimate of the win chance, or a starting point from the Probability Model if it's NFL or college football.
Check the edge. Plug that probability and the odds into the Edge Calculator. If it says no edge, stop — don't bet. If there's an edge, it tells you the Kelly-recommended stake.
Stress-test the sizing rule (optional, but worth doing once). Use Risk of Ruin with your typical edge size to see what your chosen stake fraction actually does to a bankroll over hundreds of bets, and how often it goes bust.
Place the bet, then log it in the Tracker. Settle it as a win or loss once it resolves, and let it keep an honest running record of how you're actually doing.

Tip: For NFL or college football specifically, This Week automates steps 1 and 2 — it pulls this week's games and live odds itself and shows you only the ones worth a look. You still don't need the Probability Model (or This Week) to use this app at all — with your own view on a game's win probability, skip straight to the Edge Calculator with your own number.

Home

The Command Center. Star any team or player — on their own page, or on a Trends card — and it shows up here with live context: a team's current rank and next opponent, a player's last game line. This becomes your own dashboard instead of a generic landing page. Nothing starred yet? It just shows an "Explore" grid straight to every tool below.

Research: Matchup

Everything the model sees for a specific matchup, laid out for you to dig into yourself — not to replace your own judgment, but to give it something real to work with. Open it from a Research this game link on any card in This Week (including flagged and no-edge games — an automated "no edge" call just means the model and the market agree, not that there's nothing worth a closer look), or pick any two teams yourself on the Matchup tab directly.

What's on the page

Is this a good bet?: The same pick/stake check as This Week, but for any matchup you research — not just the ones ESPN has live odds for, and not just moneyline. If you opened this page from a This Week card, the moneylines are pre-filled from the live odds already; every field is always editable, so you can check a line you found elsewhere or a purely hypothetical one. Enter a spread and/or total line (both sides default to standard -110, editable if you have the real price) and each gets its own real cover/over probability and Kelly stake, exactly like moneyline — using the actual historical spread of outcomes for that model (distribution-free, not a bell-curve assumption) rather than pretending the point estimate itself is precise. "Model implies" still shows the bare point estimate above the check — checked against real historical market lines, that estimate alone runs about 8 spread points / 4 total points off the real number on average (sometimes much more), which is exactly why the check below asks how far the line sits from the model's whole range of likely outcomes instead of trusting the single number.
Model probability, decomposed: Not just a number — how much of it came from Elo alone, versus how much the added team stats (and, for NFL, the presumed starting QBs) moved it. Lets you judge whether the model's conviction rests on one signal or several agreeing.
Elo rating trend: Both teams' last several games plotted side by side — climbing means beating expectations lately, flattening or dropping means the opposite.
Recent form & head-to-head: Plain game logs: who each team just played, what happened, and every prior meeting between these two teams in the dataset. The number next to each opponent's name is their Elo rating going into that game — a win over a 1650-rated team says a lot more than a win over a 1350 one, and Elo already accounts for that when it updates a team's own rating; this just makes that strength-of-schedule effect visible instead of only baked invisibly into the final number.
Team stat trends (NFL only): Total yards, turnover margin, and offensive efficiency, one point per game — the same numbers the NFL model actually trains on, so you can see the trend behind the diff.

Tip: Every edge card and every collapsed "flagged" or "no edge" item in This Week has its own Research link — the model only recommends a small slice of each week's games, but every game is here to look into if you have your own angle (injury news, a rivalry factor, weather) the model can't see.

Research: This Week (automated, NFL & college football)

Pick a sport, and it does steps 1 and 2 of the big picture above for you: pulls this week's upcoming games and live odds automatically, runs each one through the same probability model as the Model page (NFL includes real team-stat features; college football is Elo-only — see below), and surfaces only the games worth a second look — sized against your current bankroll from Tracker.

How to use it

Open Research → This Week, pick NFL or college football — it loads that week's schedule and odds on its own.
Pick a stake-sizing rule (defaults to ½ Kelly) if you want something other than the default.
Read the top list: each card is a specific recommendation (team, odds, expected value, recommended stake).
Click Log this bet on a card you actually place — it carries the numbers straight into Tracker's form so you don't retype them.

What the numbers mean

Model prob. / Market implied: The model's win probability for the recommended side, next to what the sportsbook's own price implies (which already includes the book's built-in margin, the "vig"). The gap between them is where the edge comes from.
Expected value: Same meaning as Expectancy on the Edge Calculator — expected profit per $1 staked, at this specific market price.
Flagged games: Games where the model disagrees with the market by a lot get pulled out of the recommendation list and shown separately instead of hidden entirely. A big disagreement with a real sportsbook's price is much more likely to mean the model is missing something (an injury, a coaching change, a scheduling quirk) than that free money was found — so these are shown as a limitation to understand, never as a bet to place.
Season so far: Each card also shows both teams' current-season points/game, yards/game, and turnover margin — just for your own read on the matchup. For college football this is pulled live from ESPN and is informational only (no historical archive exists to train on, so it isn't part of the pick or stake); very early in a season ESPN sometimes hasn't updated its own numbers yet and quietly hands back last season's final stats instead, which gets detected and relabeled "Last season (final)" so you're never looking at stale numbers labeled as current. For NFL, these numbers come from the same archive the model actually trains on — so what you see here is exactly what shaped the prediction, not a separate live read.
How would we have done?: A report card at the bottom of the page: it replays every game finished in the last 10 days as if it had been predicted beforehand — using a model trained only on games before that window, so it never saw those results — then checks the pick against what actually happened. It's the closest thing to proof this app has: not "trust the model," but "here's exactly how it did on real, already-decided games," refreshed automatically as more games finish each week.

Watch out: This is still the same simplified model from the Probability Model page — no weather or roster news factored in beyond what's described below — reading live odds from sportsbooks via The Odds API and ESPN's public, unofficial API, either of which can change or go down without notice. Treat every recommendation here as a starting point to sanity-check against what you actually know about the matchup, not a guarantee.

Research: Team & Player

Two more Research sub-tabs, for when you want to dig into one thing rather than a matchup.

Team: One team's own page: current Elo and rank, an independent second-opinion rating (SRS — margin-of-victory based, no home-field adjustment, computed entirely from final scores, never fed into the win-probability model), opponent-adjusted offense/defense rank, a rating-trend chart, recent form, and every upcoming game with the model's read on it. Tap the star next to the team name to pin it to Home.
Player: Search any player who has appeared in a box score — not just the ones with a live prop edge today. Shows season averages, the model's own projection, every book's current line, a hit-rate table (green/yellow/red, same color scale Outlier.bet uses) at last-5/last-10/season splits, a bar chart of recent games against the line, and a current streak badge. A player who only shows up as a defensive/tackle leader gets their raw game log instead — tackles aren't modeled yet, so there's no fake projection for them. Star a player here too, to pin them to Home.

Research: Games & Rankings

Games: The full schedule — not just today's picks. "This Season" shows every game played or upcoming, most current first; "History" shows every completed game from before this season, most recent first. Filter to one team, then tap any matchup to open it in Research.
Rankings: Every team, ranked by current Elo, sortable by any column — Δ Season (how far a team's rating has moved since their first game this year), SRS (the same second-opinion rating from the Team page), record, PPG, opponent-adjusted offense/ defense, and (where available) yards/game, turnover margin, and EPA/game. A link at the bottom leads to opponent-adjusted QB Rankings, the same method applied to starting quarterbacks instead of teams.

Research: Props

Player prop edges, run through the exact same de-vig + trust-guardrail pipeline as moneyline picks — a real model-vs-market edge, not just a hit-rate count. Projects each player's passing/rushing/receiving yards from their own walk-forward season-to-date average (never using games not yet played), then checks that projection against every book's posted line for a real cover probability — using the actual historical residual distribution for that stat category, not a bell-curve assumption, so a right-skewed stat (a few huge games pulling the average up) doesn't quietly bias the over/under split the way a Normal-distribution model would.

Watch out: This uses a plain season-to-date average, not an opponent-adjusted projection (that's a planned refinement) — validated separately per category on real history, and rushing, receiving, and passing all beat a naive league-average baseline by very different margins. Weight a passing or receiving edge with more caution than a rushing one. Receptions, touchdowns, and anytime-TD props aren't modeled yet — they need a different (count-based) approach than yardage does.

Trends

Situational ATS/O-U trends — deliberately different from a typical trend site: every result is measured against the real closing line (not a flat 50%, which lets "good teams cover more" masquerade as a real situational edge), and every trend shows its sample size instead of hiding it. Five sub-tabs:

Trusted Trends: A curated list (favorite/underdog splits, division games, rest, weather) — each card shows Record, ROI%, cover/over rate, Units, current Streak, Last 10, Games, and Years, plus a cumulative-units chart. ROI/Units assume a flat 1-unit bet at standard -110 on every qualifying game (there's no historical per-game price archive to do better) — a disclosed approximation, not hidden in fine print. A trend needs at least 30 decided games before it's marked trusted (a green border); most curated situations, run over their FULL history rather than cherry-picked, actually show a small negative ROI once real vig is counted — that's the honest, well-known truth that most "trends" don't beat the vig, not a bug.
Team Trends: Pick a team, and every angle worth checking against them (home/away, favorite/underdog, all four combinations, division games, rest) runs at once — the same math as Build Your Own, just pre-scoped to one team.
Today: Every team playing in a game with a posted line right now, matched to the curated trend for their exact home/away + favorite/underdog spot — so you can see what applies before kickoff without rebuilding it yourself. Needs the live-odds feed to be populated; shows nothing if it isn't.
Build Your Own: Combine your own filters (team, role, location, division game, and — NFL only — rest and weather) into a custom trend. Every combination still runs through the same 30-game trust gate and ROI/units math as the curated list, so a thin custom slice reads as "small sample," not a false edge — the usual trap with unlimited-filter trend builders, closed by construction instead of a disclaimer.
My Trends: Tap the star on any trend card (anywhere in Trends) to save it here — a personal running list, recomputed live every time you open it, not a frozen snapshot. Saved per browser, not shared with anyone else yet.

Tracker: Edge Calculator

Answers one question: given this probability and these odds, is there really an edge, and if so, how much should I stake? This is the tool to open every single time before placing a bet, for anything the rest of the app doesn't already check for you — a manual line, a different sport, a hypothetical.

What you enter

Win probability (p): Your honest estimate of how likely this bet is to win — not the odds, your own belief. If you're not confident in a precise number, err toward a more conservative (lower) estimate; overestimating p is the single most common way Kelly sizing gets people in trouble.
Net odds (b): How many units you win per unit staked if you're right. Even money is b = 1. Standard -110 odds (bet $110 to win $100) is b ≈ 0.909. For American odds: positive odds like +150 → b = 1.50; negative odds like -150 → b = 100/150 ≈ 0.667.
Bankroll: The total dollars you're managing this betting strategy with.
Stake sizing: Full / ½ / ¼ Kelly — how aggressively to size relative to the mathematically "optimal" full-Kelly stake. Defaults to ½.

What the numbers mean

Expectancy: Expected profit per $1 staked. Positive means the bet is profitable on average if you could repeat it many times; zero or negative means there's no real edge.
Std. dev: How much a single bet's outcome swings around that average — the volatility you're taking on for this expectancy.
Kelly fraction (f*): The mathematically growth-optimal fraction of your bankroll to stake, assuming your probability estimate is exactly right (it never is — that's why fractional Kelly exists).
Recommended stake: f* × your chosen fraction × bankroll, in dollars. This is the number you actually bet.
The chart: Shows how your bankroll's long-run growth rate changes as you stake more or less of it. It peaks at the Kelly point and then falls — betting more than full Kelly makes things worse, not better, and betting everything (100%) guarantees eventual ruin.

Step by step

Enter your probability estimate and the odds being offered.
Enter your current bankroll.
Check the top of the page — if it says "no edge, do not bet," stop here regardless of how good the bet feels.
If there is an edge, read off the recommended stake at ½-Kelly (the default) and bet that amount.

Watch out: A "55% edge" that's actually 50.5% will still get sized aggressively by full Kelly — the formula has no idea your probability estimate might be wrong. That's the entire reason to default to fractional Kelly rather than full.

Tracker: Risk of Ruin

Answers: if I keep betting this way for a long time, what actually happens to my bankroll — and how does that differ from what the "average" outcome makes it look like? Use this occasionally to sanity-check a sizing habit, not before every bet.

What you enter

Win probability / odds: The typical edge you expect to find repeatedly.
Bets per bettor (M): How many sequential bets to simulate — e.g. a season's worth.
Simulated bettors (N): How many independent, parallel copies of "you" to simulate. More bettors = a smoother, more reliable ruin-rate estimate.

What the numbers mean

Ruin rate: The % of simulated bettors whose bankroll ever collapsed to ≤1% of where they started.
Ensemble average vs. median: The average is dragged upward by a few enormously lucky paths — it is not what a typical bettor experiences. The median is the typical single path. The "avg ÷ median" stat shows exactly how misleading the average is; it's usually well above 1×.
The fan chart: Faint lines are individual simulated bankrolls (log scale, so both huge gains and near-ruin are visible on the same chart). The orange line is the average; the blue line is the median. The gap between them is the lesson.

Step by step

Set p/b to match the kind of edge you typically find.
Click through Full Kelly → ½ Kelly → Flat → 2× Kelly and compare the ruin-rate bars at the bottom.
Pick the sizing rule with a ruin rate you're actually comfortable with — not just the one with the highest average outcome.

Tip: If 2× Kelly's ruin rate looks shocking compared to full Kelly's, that's intentional — overbetting even a real edge is one of the fastest ways to go broke, and this chart is the clearest way to see why.

Tracker: Bets & CLV

Your actual betting log. Answers: am I really making money, and if not, is it my picks or my sizing that's the problem?

How to use it

Set your starting bankroll once, at the top.
Every time you place a real bet, log it: description, the p you used, the odds, and your actual stake. It starts as "pending."
Once the game/event resolves, click Win or Loss on that row to settle it — or wait, since a bet logged from a This Week/Research/Trends card carries enough structured data to auto-settle itself against the final score the next time you open this page.
Watch the stats and chart update — that's your real track record, building up over time.

What the numbers mean

Current bankroll / Total P&L: Where you actually stand right now, in dollars.
Win rate: % of settled bets you've won (pushes don't count against you).
Current / max drawdown: How far below your peak bankroll you are right now, and the worst it's ever been. Large drawdowns are normal even with a real edge — that's the volatility from Risk of Ruin showing up in real life.
Realized vs. theoretical (the chart): "Realized" is what actually happened with your real stakes. "Theoretical" replays the exact same bets and outcomes, but sized at recommended fractional-Kelly instead. If realized tracks noticeably below theoretical, your picks are fine but your sizing is costing you money — bet closer to the recommended amount. If it's above, you've been sizing more aggressively than recommended, which means better luck so far but more risk than the model would advise.
Closing line value (CLV): Once a logged bet's game reaches kickoff, shows whether the price you actually got beat the closing line — the standard skill signal sharp bettors track independent of whether any individual bet won, since a good price beaten by bad luck is still a good decision.

Watch out: This data lives only in this browser's local storage — it is not backed up or synced anywhere. Clearing your browser data or switching devices/browsers loses it.

Guide: Probability Model

Where the probability in step 1 can come from instead of a gut estimate, for NFL or college football specifically. It trains a small, transparent model on real historical games and is upfront about how much to trust it. Lives here in the Guide rather than its own top-level nav item, since it's methodology, not a tool you visit weekly.

How to read it

Dataset: Pick NFL or college football. Everything below recomputes for that sport.
Elo K-factor / home advantage: Tuning knobs for the underlying team-strength ratings. Higher K-factor makes ratings react faster to recent results; you can usually leave these at their defaults.
Season regression: Pulls every team's rating partway back toward average the first time it plays in a new season, to account for roster turnover between seasons — especially relevant for college football, where rosters change far more than the pros. Defaults to 35%, which measured best overall on real history. It is a general accuracy tune, not specifically an opening-week fix — the very earliest games of a season actually calibrate a touch worse as this goes up, even though the season overall gets more accurate.
"vs. Elo-only" (NFL only): On NFL, the model also factors in each team's season-to-date total yards, turnover margin, and offensive efficiency — using only games already played before the one being predicted, never a peek ahead. This tile trains the same model with those extra stats turned off and compares the two head-to-head, so you can see the real effect instead of just trusting it helps. College football doesn't have a public per-game stats archive like this, so its model still runs on Elo alone.
Brier score: The model's calibration grade — lower is better, 0 is perfect. What matters more than the raw number is the comparison right next to it: does it actually beat a naive "always guess the average home-win rate" baseline? If not, the model isn't adding value.
Reliability diagram: The single most important chart on this page. Each dot groups games by the model's predicted probability and plots the probability it actually assigned against how often that side actually won. Dots sitting on the dashed diagonal mean the model is honest — when it says 70%, that side wins about 70% of the time. Dots below the diagonal mean the model is overconfident, which is the dangerous direction for Kelly sizing.
Backtest: Replays only the held-out games (the ones the model never trained on) as bets at standard -110 odds, comparing flat vs. full-Kelly vs. half-Kelly sizing.
Per-bet growth rate: The number to actually compare across strategies — not the raw final-bankroll dollar figure, which can compound into meaningless, astronomically large numbers over 1,000+ backtested bets (see the warning below).

Step by step

Pick a sport, and check the Brier score beats the naive baseline.
Look at the reliability diagram — if dots track the diagonal reasonably well, the model's probabilities are trustworthy enough to use as a starting point.
Compare per-bet growth rate across flat / full-Kelly / half-Kelly in the backtest to get a feel for how sizing discipline would have mattered historically.
For a real upcoming game, take the model's implied probability as a starting point — not gospel — and carry it into the Edge Calculator, ideally shaded a little toward 50% to account for the model's own uncertainty.

Watch out: The backtest's full-Kelly dollar figure can look absurd (far beyond anything real). That's expected — it's genuine Kelly math compounding a sustained edge across 1,000+ consecutive bets, assuming a market that would let you keep re-betting an ever-growing bankroll at unchanged odds for over a decade, which no sportsbook would actually allow. Treat it as an illustration of how powerful (and dangerous) uncapped compounding is, not a real projection — and use the per-bet growth rate instead.

Key ideas to remember

No edge, no bet — always check expectancy before staking anything.
The average is not what you get — your own single path looks like the median, not the ensemble average. Judge sizing rules by ruin rate and the median outcome, not the average.
Fractional Kelly is the default, not a compromise — your probability estimate always has error, and Kelly has no idea how much.
A calibrated model beats a confident one — an overconfident model is more dangerous to size bets off of than no model at all.
Sample size before conclusions — a trend with 100% win rate over 12 games and a trend with 55% over 500 games are not equally trustworthy; Trends flags this everywhere on purpose.
Track sizing and picking separately — Tracker's realized-vs-theoretical comparison tells you which one is actually costing you money.