1. Predicting Substitutions
in the NBA
Charlie DeStefano, Ph.D.
Consulting for Unblockable, Inc. (CTO & Game Designer)
HotStreak
2. Motivation
Enormous Sports Betting Market
• $107B currently spent annually in U.S.
• $148B with legalization
Provide short-term, player proposition bets
• Substitutions not currently considered
• ~30% of all bets suffered from early substitutions
Goals:
1. Predict when substitutions will occur
2. Adjust bet lines and payouts accordingly
3. Solution
Analyze historical substitution patterns (Past 2
seasons)
1. Regression
• Predict time until next substitution
• Gradient Boosting Regression
2. Survival Analysis
• Predict probability of remaining on-court until a
substitution occurs, based on past observations
from similar scenarios
Deliverable: Updated
Bet Offerings
4. LeBron James:
• Current Time: 5:00 remaining
• Over/Under: 4 points
• End Bet Time: 0:00 remaining
• Next Sub Prediction 1:14 remaining (75% shorter)
• Survival Probability 61%
Updated Bet
1. New Line set to 3 points (.75x4)
2. “Under” Payout Reduced by 39% (100--61)
3. “Over” Payout Increased by 39%
Example Scenario
6. Pipeline
Raw Game, Player,
& Bet Data
Feature
Engineering
Calculate
Regression
& Survival
Models
Compress
Models for
Deployment
Acquire
Real-Time
Game Data
CSV Files Pandas
SKLearn &
Lifelines
Pickle
Licensed Data
Sources
Calculate New
Bet Lines &
Payouts
Offer New
Bet to
User
Real-Time, During
Games
Pre-Processed Before
Games
7. Historical Bet Analysis
$1 Bets
Original Lines
and Payouts
New Lines and
Payouts
Player Wins 53,802 53,832
Player Losses 53,924 53,894
Paid Out $53,802 $45,916
Paid In $53,924 $53,894
Profit $122 $7,978
• 107,726 total bets
• 71% of all wins were “Under” bets
• Majority of wins were “Under” bets =
Majority of Payouts reduced