Excited to share a 4th down decision bot that models how each SEC coach behaves on 4th down. Took me a second to publish with memory issues on the free tier of R Shiny 😂.App: Code:

RT @ShamsCharania: BREAKING: San Antonio Spurs star De'Aaron Fox has agreed to a four-year, $229 million maximum contract extension with th….

Example of the App Output:

RT @tejfbanalytics: never change lane kiffin. our 4th down go for it goat

Appreciate any feedback, DMs open and would love to connect!.

Limitations + next steps:.– App crashes if you change team + yardage in one go, please make sure to reload before each change.– Opponent strength, weather, and injuries not factored in.– WP model uses average tempo, not team-specific style.– Data is only including 2019-2024.

Frontend:.– Custom field layout with team logos and fonts.– Inputs for clock, field position, score, etc. – Outputs coach’s likely decision + EPA and WP for all 3 options.– Lists top 50 most similar historical plays.Backend:.– Models served via Dockerized Google Cloud Run APIs.

Model accuracy:.– Coach classifiers: 0.81–0.91 accuracy, kappa up to 0.86.– Win probability model (global):.• RMSE: 0.032.• MAE: 0.015.• R²: 0.991.All models trained using tidymodels, with coach classifiers using XGBoost and WP using random forest.

Used cfbfastR 4th down data from 2019–2024. Each model is coach-specific, trained only on plays from that coach’s history. For example: Texas uses plays from Sark at Texas + Alabama while Alabama uses plays from DeBoer at Fresno State, Washington, and Alabama.

Key features:.– Trained a separate ML model for each SEC team using their historical 4th down decisions.– Models predict whether a coach would go for it, punt, or kick a field goal.– Then simulate each option to calculate expected points and win probability.

@WillStokesV Limitations & feedback: .– Model assumes normal error using RMSE from 2021–2024 drafts .– Doesn’t account for team fit, positional needs, or trades .– Limited to Day 1: players outside Round 1 are assigned pick 33 .Open to feedback or suggestions to improve!.

@WillStokesV Cont’d:.– UI compares 2025 analyst picks to actual results with avg error .– Added headshots + logos for visual clarity .– RMSE calculated on 2021–2024 data to estimate uncertainty.

@WillStokesV Notes:.– Used mock data from Daniel Jeremiah, Mel Kiper, Bucky Brooks, Mike Tannenbaum, Josh Norris, and Consensus boards (2021–2025) .– Built regression model to predict actual draft position .– Visualized availability curves + pick probabilities.

Finally posting a Shiny app I’d built a couple months ago that estimates the probability a player will be available at your team’s pick in the 2025 NFL Draft (worked with @WillStokesV on this). ShinyApp: Code:

RT @avsportsanalyst: @VarunRaman2205

RT @avsportsanalyst: @VarunRaman2205 The lowest prospect is set at 30 and every other QB is relative to that.

RT @avsportsanalyst: @VarunRaman2205 PCA was used to create a metric for each season based on the 4 statistics described previously. The we….

RT @avsportsanalyst: @VarunRaman2205 The EPA/play, ANY/A, success rate, and SoS of the team were used. They were correlated to the draft pi….

RT @avsportsanalyst: @VarunRaman2205 The quarterbacks were filtered so that they have at least 250 pass attempts in a certain CFB season an….

RT @avsportsanalyst: Along with @VarunRaman2205, we created an "Analytical Big Board" rating for 2024 NFL Draft QB prospects. Methodology….