-
Optimize Full Season with Exponential Value Decay (Temp. API; full version coming soon) |
Optimize 10 Weeks, No Value Decay (Temp. API; full version coming soon) |
Source Code
-
Suicide pools are a common and very fun office pool in which you are required to pick a winner of an NFL game each week. If your pick is right (straight up / no spread or odds), you advance. If it is wrong, you're eliminated. Last man or woman standing wins. For this tool I've built in the following standard rules:
- One pick per week
- Single elimination with no "buying back in"
- You cannot pick the same winning team twice in the season
-
Since its very unlikely for a player to survive for the full season, the objective is to maximize the probability that you outlast all opponents in your pool. The key is to use win probabilities to survive early, but also to look at future matchups to ensure good teams are saved when they will be needed in the future. To do so, it implements a few simple tools:
- Pairwise match-up win probabilities, implied from 538.com's published FPI rankings
- Informed random samples of winners each week
From each of these paths, the model updates and begins to converge by identifying team-week pairs that very regularly result it high-value paths. While it is parameterized to try to keep the chances of a global maxima high, it may converge to a local max.
The naive/simple implementation would be to simply optimize for your survival over the full season, however this can result in saving good teams for future weeks that may be valueless (if all opponents are eliminated) while lowering your survival chances this week. As such, the tool is built to optimize either linearly or with an exponential decay over either a fixed number of weeks or the full remaining season.
