Using nflendzoneModel Joint NegBinom Workflow

View repository source

This draft is focused on one path: fitting team_strength_bivar_negbinom and producing the team_strength_negbinom_summary artifact consumed downstream.

End-To-End Flow

Show code

flowchart LR
  A[nflverse schedule + results] --> B[prepare_schedule_indices]
  B --> C[prepare_stan_data]
  C --> D[fit_team_strength_model model=team_strength_bivar_negbinom]
  D --> E[fit$summary and rvar extraction]
  E --> F[team_strength_negbinom_summary release]
  F --> G[nflendzoneApp probability reconstruction]

flowchart LR
  A[nflverse schedule + results] --> B[prepare_schedule_indices]
  B --> C[prepare_stan_data]
  C --> D[fit_team_strength_model model=team_strength_bivar_negbinom]
  D --> E[fit$summary and rvar extraction]
  E --> F[team_strength_negbinom_summary release]
  F --> G[nflendzoneApp probability reconstruction]

Minimal Fit Pattern

This mirrors inst/examples/update_negbinom_team_strength.r.

Show code

fit_stan_data <- prepare_stan_data(
  game_data = training_data,
  teams = teams,
  verbose = TRUE
)

fit_bivar_negbinom <- fit_team_strength_model(
  stan_data = fit_stan_data,
  model = "team_strength_bivar_negbinom",
  seed = 52,
  chains = 4,
  parallel_chains = parallel::detectCores(),
  iter_warmup = 1000,
  iter_sampling = 1000,
  adapt_delta = 0.95,
  max_treedepth = 10
)

Build The Summary Artifact

Show code

week_vars <- c(
  "phi_home",
  "phi_away",
  "filtered_alpha_log",
  "filtered_team_off_strength[team]",
  "filtered_team_def_strength[team]",
  "filtered_team_hfa[team]",
  "filtered_league_hfa",
  "predicted_alpha_log",
  "predicted_team_off_strength[team]",
  "predicted_team_def_strength[team]",
  "predicted_team_hfa[team]",
  "predicted_league_hfa"
)

nb_sum <- fit_bivar_negbinom$summary(
  variables = stringr::str_extract(week_vars, "^\\w+")
)

Publish For App Consumption

Show code

pb_write(
  x = nb_sum,
  file = paste0("team_strength_negbinom_summary_", filter_season, ".rds"),
  repo = github_data_repo,
  tag = "team_strength_negbinom_summary"
)

Why This Model Path Matters

Joint home/away count modeling is more natural for score-like outcomes than point-difference-only views.
The release summary can be compact while still enabling stochastic reconstruction in the app.
The same artifact supports weekly diagnostics, slate previews, and betting-probability calculations.

Practical Checks Before Shipping

Validate diagnostics (rhat, ESS, divergences) before publishing.
Ensure season/week identifiers in the summary align with release timestamp metadata.
Compare implied game-level probabilities against previous week to catch structural jumps.

If this post is useful, the evergreen system docs live in Projects Architecture and package docs live in Projects Packages.

--- title: "Using nflendzoneModel Joint NegBinom Workflow" description: "Draft guide to the joint negative-binomial team-strength workflow in nflendzoneModel, from Stan data prep to app-facing summary artifacts." author: Tyler Pollard date: 2026-02-15 date-modified: 2026-02-15 draft: false categories: - nflendzone - nflverse - cmdstanr - brms - stan - bayesian - negbinom image: /logos/tyler-pollard-icon.png repo-url: "https://github.com/TylerPollard410/nflendzoneModel" --- [View repository source {{< bi github >}}]({{< meta repo-url >}}){.btn .btn-secondary target="_blank"} This draft is focused on one path: fitting `team_strength_bivar_negbinom` and producing the `team_strength_negbinom_summary` artifact consumed downstream. ## End-To-End Flow ```{mermaid} flowchart LR A[nflverse schedule + results] --> B[prepare_schedule_indices] B --> C[prepare_stan_data] C --> D[fit_team_strength_model model=team_strength_bivar_negbinom] D --> E[fit$summary and rvar extraction] E --> F[team_strength_negbinom_summary release] F --> G[nflendzoneApp probability reconstruction] ``` ## Minimal Fit Pattern This mirrors `inst/examples/update_negbinom_team_strength.r`. ```{r} #| eval: false fit_stan_data <- prepare_stan_data( game_data = training_data, teams = teams, verbose = TRUE ) fit_bivar_negbinom <- fit_team_strength_model( stan_data = fit_stan_data, model = "team_strength_bivar_negbinom", seed = 52, chains = 4, parallel_chains = parallel::detectCores(), iter_warmup = 1000, iter_sampling = 1000, adapt_delta = 0.95, max_treedepth = 10 ) ``` ## Build The Summary Artifact ```{r} #| eval: false week_vars <- c( "phi_home", "phi_away", "filtered_alpha_log", "filtered_team_off_strength[team]", "filtered_team_def_strength[team]", "filtered_team_hfa[team]", "filtered_league_hfa", "predicted_alpha_log", "predicted_team_off_strength[team]", "predicted_team_def_strength[team]", "predicted_team_hfa[team]", "predicted_league_hfa" ) nb_sum <- fit_bivar_negbinom$summary( variables = stringr::str_extract(week_vars, "^\\w+") ) ``` ## Publish For App Consumption ```{r} #| eval: false pb_write( x = nb_sum, file = paste0("team_strength_negbinom_summary_", filter_season, ".rds"), repo = github_data_repo, tag = "team_strength_negbinom_summary" ) ``` ## Why This Model Path Matters - Joint home/away count modeling is more natural for score-like outcomes than point-difference-only views. - The release summary can be compact while still enabling stochastic reconstruction in the app. - The same artifact supports weekly diagnostics, slate previews, and betting-probability calculations. ## Practical Checks Before Shipping 1. Validate diagnostics (`rhat`, ESS, divergences) before publishing. 2. Ensure season/week identifiers in the summary align with release timestamp metadata. 3. Compare implied game-level probabilities against previous week to catch structural jumps. {{< include ../../../_footer.qmd >}}