About & Methodology
An NFL model that shows its work.
Most prediction sites tell you who to pick. GameLens tells you why, how confident it is, and how it has done when it was wrong.
Who built GameLens, and why
GameLens is built by Aden Khan, a freshman at the University of Miami who is originally from the Bay Area. Aden loves sports betting, deep analytics and the game of football as a whole, and GameLens is where those three things meet.
The idea was simple: build an NFL product that explains its reasoning, lets fans explore scenarios, and publicly evaluates its own predictions. Too much football content is either box scores that describe what happened without explaining why, or picks sold with no track record attached. GameLens tries to be the opposite: it describes each team's identity in plain language, lets you simulate the rest of the season yourself, and publishes every prediction next to its result.
GameLens is an independent project. It is not affiliated with the NFL, ESPN, or any sportsbook, and it does not sell picks.
For how it was built, the architecture, and an honest look at results and limitations, read the project case study.
Data sources
Schedules, scores, play-by-play, box scores, rosters, injuries and betting lines come from ESPN's public NFL data feeds. The feed does not supply expected points; GameLens computes its own approximate EPA from each play's down, distance and field position (see below). Team logos and player headshots are served from ESPN's image CDN.
Everything else — team ratings, scheme and coverage profiles, Impact Ratings, usage trends, prop probabilities and season simulations — is computed by GameLens from that raw data.
How GameLens EPA is calculated
Expected points (EP) before a snap are estimated as 6.4 − 0.072 × yards to the end zone, minus a down penalty (0, 0.55, 1.25 and 2.1 points for 1st through 4th down) and a distance adjustment relative to a typical distance to gain, clamped between −2.5 and 6.8. A play's EPA is the EP of the next scrimmage snap minus the EP before it (from the other team's perspective after a change of possession), with touchdowns valued at a fixed 7 points and each play clipped to ±10.
This is an approximation chosen to resemble published expected-points curves. It is not equivalent to nflfastR or any fitted model, and it has not been validated against one. Known limitations: it ignores time remaining, score, timeouts and the half; it skips past special-teams plays to find the next snap, so punts, field goals and kick returns are not valued directly; it treats every touchdown as 7 points regardless of the conversion; and defensive scores, turnovers on downs, safeties and half-ending plays are handled coarsely. Treat play-level values as directional and season aggregates as more reliable than single plays. Changes to this formula ship with a new model version.
Model philosophy
Efficiency over outcomes. Points and wins are noisy. A team's EPA per play stabilizes faster and predicts next week better than its record, so ratings lean on efficiency.
Respect the market. Betting lines aggregate a huge amount of information. Rather than ignoring them, the model starts from the opening line and moves away from it only as far as its own evidence justifies.
Shrink small samples. Three games is not a season. Every rating, projection and probability is pulled toward average until the sample earns its way out, which is why early-season edges are deliberately modest.
Grade honestly. A model that is only shown when it is right is marketing. GameLens tracks straight-up record, against-the-spread record, and closing line value for every flagged pick.
How predictions are generated
Grade every play
Each scrimmage play is scored by GameLens’s approximate expected points added (EPA): a hand-built field-position, down and distance formula, not a trained expected-points model.
Rate every team
Offensive and defensive EPA per play, relative to league average, blended 55/45 with point margin and shrunk toward zero for small samples.
Project the game
The rating gap plus 1.7 points of home field gives a stats-only margin and total.
Blend with the market
The opening line is treated as a prior. The model earns between 25% and 70% of the weight as the season sample grows.
Convert and flag
The margin becomes a win probability. A spread pick is only flagged when the model disagrees with the line by at least a point.
Player props use the same principles: a projection from each player's rate stats and recent role, pulled toward the posted line by sample size, then converted to a capped over/under probability. The season simulator plays out the remaining schedule 5,000 times using the same team ratings.
How often data updates
- Live games: scores, win probability and play-by-play refresh about every 15 seconds.
- Injury reports: refreshed every 15 minutes.
- Ratings and the slate: rebuilt in the background every 30 seconds from the latest completed games.
- Final games: locked once official and cached for an hour, since they no longer change.
- Personnel and depth charts: refreshed every six hours.
- New week: the slate rolls over at 3:00 AM Pacific on Tuesday.
What “out of sample” means
A prediction is out of sample when the model making it had never seen the result. It sounds obvious, but many published records quietly break this rule by grading old games with ratings that already include those games.
GameLens rebuilds its ratings for each week using only games completed before that week kicked off. When the Model page says the Week 6 picks went 10-4, those were the picks a Week 6 model would actually have made. That is the only kind of record worth trusting.
Recorded vs. reconstructed. Since Week 3, GameLens saves each pick to a timestamped archive before kickoff, along with the model version, the input cutoff and the market line at that moment. Once a game kicks off, that saved pick is frozen and is the one shown everywhere on the site; only the final score is added afterward. You can review it in the prediction archive. Records for earlier weeks, and the season-long backtest and CLV figures, are still reconstructions: they re-run each week's pre-kickoff data and show what the method would have done, not picks posted in advance.
What GameLens is trying to accomplish
Make advanced football analytics understandable without dumbing it down. Every number on the site should answer a real question — who is better, who is being used more, how likely is this — and every unfamiliar term has a plain-English definition one hover away.
GameLens is for analysis and entertainment. Probabilities are not guarantees, and nothing here is betting advice.