Welcome back to the series! In the previous installment, we dissected the strategies for Week 4 of the 2023 NFL season using k-means clustering. If you missed it, you could catch up here. This week we made major upgrades to the analysis.
Methodology (In-Brief)
- Aggregated projections from six sources for the first five weeks
- Collected the actual results of the first four weeks.
- Collected information about the slate including:
- Projected Spread: The expected point difference between the two teams set by Vegas.
- Game Total: The total expected points to be scored in the game.
- Matchup Rating: Based on the average fantasy points allowed to a player’s position by the opposing team.
- Salary: We will focus on the DraftKings Main Slate in these.
- Value: The standard DFS Metric (projection/salary) * 1000
- Risk
- Reward
- Uncertainty
- Filtered out only the top 250 players to reduce the player pool.
- Split out all the players by position.
- Performed a factor analysis and created a scree plot for each position to determine the appropriate number of clusters.
- Performed a K-Means Cluster Analysis on each position. (If you are unfamiliar with this machine learning technique, I would recommend StatQuest on YouTube for an excellent breakdown.)
- Pivoted the Data for visualization.
Clustering Results
This week I remapped the clusters for easier use, so Cluster 1 is the best and higher numbers are worse. They are sorted by actual multiplier.
This week I also added a 2x and 3x column to the viz which helps a ton with GPPs in particular.
For example, Cluster 4 has the fourth worst average multiplier, but of the WRs to hit at least 3.0x so far this season, 26.32% came from this group.
Using a similar logic, the GPP specific clusters are:
- QB: 2
- WR: 4
- RB: No clear winner here.
- TE: 3
- DST: 3 and 4
Week 5 Player Pool
Rather than speculate about which specific players you should start, I will instead focus on three tasks unique to this data analysis:
- Reduce the player pool as much as possible to let you focus your research
- Identify players who look like traps. These are players with high projected values, yet something about their combination of price, risk, and reward puts them into a cluster that has underperformed on average so far. This week, Wan’Dale Robinson is a great example.
- Identify tournament plays with an analytical approach other than strictly projected ownership
