NBA Analysis

Overview

This is a project for NBA analysis built using the Kedro package.

Pipeline overview:

dp

This is a data processing pipeline to extract data from basketball reference.

Steps involved:

• Get season data from basketball reference
  • Multiple seasons are collected with different nodes
• Clean data and merge into single file

Todo:

• Can we output a partitioned parquet dataset partitioned by season year?

shooting_per

This pipeline estimates the shooting percent of players and produces a plot of a random subset of players.

Steps involved:

• Build bayesian model to get posterior distributions for player scoring %
  • Initially start with all players have uninformed beta distributions
  • This prior is taken from the parameters.yml
• Take a random subset of players and plot into a single figure
  • The number of players to plot is from parameters.yml

Typical results:

team_scores

TODO A model that describes each team's attack and defensive abilities.

Model on score differences. Fit with gradient descent in a sequential manner. Probably with pytorch. Use the fitted model to simulate future games scores.

Docker

Create docker image from our kedro project

kedro docker init
kedro docker build

Test run:

docker run nba-analysis

Convert to local tar file

docker save -o nba_analysis_docker.tar nba-analysis

Load on another machine

docker load -i nba_analysis_docker.tar

Todo

• Converge with data sequentially
  • Give new posterior after each season
  • match by match
• Hierarchical model which gives position based priors
• Extend so that each player has a different ability per year
• Build dashboard to explore different players
• team_scores pipeline
• Github CI/CD
• Push project to google app engine on completion
• unit tests
• docs/mkdocs