How the Algorithm Works

A transparent look at the 11-factor weighted composite model behind every ClutchPicks prediction

Overview

ClutchPicks uses an 11-factor weighted composite algorithm to generate NBA spread, moneyline, and over/under predictions for every game. Instead of relying on a single data point, the algorithm synthesizes multiple independent signals — recent performance, player stats, injuries, situational advantages, market pricing, and crowd wisdom — into a unified probability estimate.

The model runs entirely in your browser using real-time data from ESPN, sportsbook odds, and prediction markets. No picks are pre-selected by humans; every prediction is algorithmically generated based on the current data.

The 11 Factors

1. Recent Form (Last 10 Games)
15%
Source: ESPN Schedule API
We analyze each team's last 10 completed games to calculate win percentage, average points scored, average points allowed, and net rating. Teams on hot streaks or in slumps are weighted accordingly. A team's net rating (points scored minus points allowed) over the last 10 games is one of the strongest predictors of near-term performance.
2. Home/Away Splits
10%
Source: ESPN Schedule API
NBA teams perform measurably differently at home versus on the road. We calculate each team's home win percentage and road win percentage separately, then factor in a 3% baseline home court advantage — the league-wide average since 2019. Teams with extreme home/road splits (like some altitude-affected teams) get appropriately amplified signals.
3. Head-to-Head Record
5%
Source: ESPN Schedule API
How have these specific teams fared against each other this season? We look at the season series record and average margin of victory. While small sample sizes limit this factor's weight, it captures stylistic matchup advantages that aren't visible in overall records — like how a switch-heavy defense might consistently neutralize a pick-and-roll-heavy offense.
4. Schedule Fatigue
10%
Source: ESPN Schedule API + Arena Coordinates
Goes beyond simple rest days. We detect back-to-backs, calculate games played in the last 72 hours, compute travel distance between arenas using GPS coordinates for all 30 NBA venues, and apply timezone-crossing penalties. A team on a back-to-back after traveling 2,000+ miles faces compounding fatigue. Penalties: B2B (-0.15), 3 games in 72h (-0.10), 2,000+ mi travel (-0.10), 1,000+ mi (-0.05), timezone crossings (-0.05).
5. Pace of Play
8%
Source: ESPN Schedule API
This factor primarily drives over/under predictions. We calculate each team's average total points (combined score) from recent games. When two up-tempo teams meet, the predicted total rises; when two grind-it-out defensive teams clash, it drops. The predicted total is compared against the sportsbook's posted line to identify over/under value.
6. Sportsbook Odds
17%
Source: The Odds API (DraftKings, FanDuel, BetMGM, Caesars)
Sportsbook lines reflect the collective wisdom of professional oddsmakers and millions of dollars in betting action. We convert American odds to implied probabilities and use the book's spread and total as anchor points. Rather than ignoring market pricing, we treat it as a valuable input — then look for where our other factors disagree. Lines from DraftKings and FanDuel are displayed side by side in every game card.
7. Prediction Markets
5%
Source: Polymarket API
Prediction markets aggregate crowd-sourced probabilities from thousands of participants who have real money at stake. When available, we pull win probabilities from Polymarket's NBA game markets. This provides an independent probability estimate that sometimes diverges from sportsbook lines, offering additional signal.
8. Star Player Impact
12%
Source: ESPN Roster + Player Stats APIs
The second-heaviest factor. We identify each team's top players by PPG and usage rate, calculate a composite "star power" score, then adjust for injuries. When a team's leading scorer is OUT, their star power drops significantly — a 28 PPG player sitting out can swing a game by 5+ points. This factor captures what the old model missed: the massive impact of individual stars on game outcomes.
9. Injury Adjustment
8%
Source: ESPN Injuries API
Beyond star players, we assess the cumulative impact of ALL injured, out, and questionable players on each roster. Multiple role players being out can matter as much as a single star — it reduces rotation depth, forces mismatches, and increases fatigue for remaining players. Each injury status (OUT, Doubtful, Questionable, Probable) carries a different impact weight.
10. Offensive/Defensive Efficiency
7%
Source: ESPN Team Statistics API
Net Rating (Offensive Rating minus Defensive Rating) is one of the most predictive team-level statistics in basketball analytics. A team that scores efficiently and defends well has a sustainable advantage that raw win-loss records can mask. We use season-long net rating as a more stable predictor than recent form alone, capturing true team quality.
11. Player Matchup
3%
Source: ESPN Roster + Player Stats APIs
A lightweight positional comparison: we match each team's starters by position (PG vs PG, C vs C, etc.) and compare their scoring averages and efficiency. While NBA basketball is too fluid for rigid positional matchups, this factor provides a small signal about which team has the edge in individual talent across the starting lineup.

From Factors to Picks

1
Score each factor — Every factor produces a normalized score between -1 and +1 (positive = favors home team, negative = favors away team).
↓
2
Weighted composite — Multiply each factor score by its weight (17%, 15%, 12%, etc.) and sum to get a single composite score.
↓
3
Logistic transform — Pass the composite through a logistic (sigmoid) function to convert it into a win probability between 0% and 100%.
↓
4
Generate predicted spread — Convert win probability to a predicted point spread. A 60% win probability roughly translates to a 3-point spread.
↓
5
Compare to book lines — Calculate the "edge" (difference between our predicted spread and the sportsbook line). Larger edges suggest potential value.
↓
6
Generate explanation — The top 2-3 factors are ranked by impact and translated into natural language explanations for each pick.
↓
7
Assign confidence — Confidence ratings (50-95%) reflect both the size of the edge and the agreement between independent factors. When form, odds, player data, and markets all point the same direction, confidence is high.

Pick Types

Spread Picks

We compare our predicted spread to the sportsbook's posted spread. If we predict the Knicks winning by 5 but the book has them at -3.5, we pick NYK -3.5 with the edge being the 1.5-point difference. Picks show actual American odds from the book, e.g., "NYK -3.5 (-110)".

Moneyline Picks

Our win probability is compared to the sportsbook's implied probability. If we give a team a 65% chance of winning but the book implies only 55%, there's value on that team's moneyline. Picks display as "DET ML (+150)" with the actual odds number prominently shown.

Over/Under Picks

Using pace analysis (Factor 5), we estimate the combined total points for each game and compare it to the sportsbook's posted total. Two up-tempo teams facing each other might have a predicted total 5+ points above the book's line, triggering an OVER pick with high confidence.

+EV Engine

Every pick now includes Expected Value (EV) calculations. EV measures how much you'd expect to profit per $100 wagered, based on the difference between our probability estimate and the sportsbook's implied probability.

EV Formula: EV = (ourProb x profit) - ((1 - ourProb) x $100)

Positive EV (+EV) means you'd expect to profit long-term at those odds. The +EV Plays section surfaces the day's best positive-EV opportunities across all games and markets.

Kelly Criterion

For +EV picks, we calculate optimal bet sizing using the Quarter-Kelly criterion: f = (b*p - q) / b * 0.25, capped at 25% of bankroll. Quarter-Kelly provides a conservative approach that reduces variance while still capturing edge.

Best-Line Finder

We scan odds from DraftKings, FanDuel, BetMGM, and Caesars to find which book offers the best price for each pick. Even small differences in odds (e.g., -110 vs -105) compound significantly over time.

Usage Vacuum

When a star player is ruled OUT, their production doesn't disappear — it redistributes. Our Usage Vacuum engine calculates how scoring, rebounding, and assists flow to remaining active players using a 70% recapture rate. Player mini-cards show adjusted projections when injuries are present.

Poisson Projections

Player scoring projections use Poisson distribution modeling to estimate the probability of hitting specific stat lines. For example, "O 24.5 pts: 62%" means there's a 62% probability a player exceeds that threshold based on their adjusted scoring mean.

Best Bets vs Value Plays

Best Bets are simply the highest-confidence picks of the day — games where multiple factors align strongly in one direction. They represent our most confident calls.

Value Plays highlight games where our algorithm disagrees most with the market (3+ point edge). These aren't necessarily our most confident picks overall, but they represent the largest discrepancies between our model and the sportsbook line — potential mispricing that sharp bettors look for.

"Why This Pick" Explanations

Every pick includes a brief natural language explanation of the top factors driving the recommendation. The algorithm ranks all 11 factors by their weighted impact on each game, selects the 2-3 most influential ones, and generates a sentence explaining each. For example:

"Boston is 8-2 in their L10 with a +9.3 net rating. Dallas is missing Luka Doncic (calf), removing 28 PPG from their lineup."

Data Loading: 3-Tier Progressive Strategy

To keep the experience fast, data loads in three progressive tiers:

Data Freshness

Limitations

No model is perfect. Our algorithm does not account for:

These limitations are why we always emphasize: these picks are for entertainment and informational purposes only. Use them as one input alongside your own research, and always gamble responsibly.