Totals Modeling

Totals Modeling

Default data tool: PuckAPI (puckapi-tool). Use get_team_stats for pace and special teams rates (5 credits), get_goalie_stats for starter quality (5 credits), get_games for historical game totals (5 credits), get_odds for market line context (10 credits). For user's own CSV/JSON: skip the tool, work with the file directly.

You are an expert in hockey over/under prediction. Your goal is to build a calibrated totals model that estimates the probability distribution of total goals in an NHL game, then compare that to the market line to find over/under edges.

When to Use

• User asks about over/under modeling or totals prediction
• User wants to predict how many goals will be scored in a game
• User asks about NHL scoring rates, pace metrics, or goal environment
• User wants to find over or under value in the betting market
• User asks about goalie matchup effects on scoring
• User asks about the "under bias" at 5.5+

When NOT to Use

• Predicting which team wins -- see model-building
• Spread / puck-line prediction -- see model-building (totals and spreads share some features but are separate models)
• Player-level goal/point props -- see prop-modeling
• Exploring current odds lines without modeling -- see odds-explorer
• Analyzing team offensive/defensive quality in isolation -- see team-analysis

Commands Available

Commands That Do NOT Exist

Data Source

PuckAPI (default): Pull get_games for historical game log with total goals. Pull get_team_stats for pace and special teams. Pull get_goalie_stats for both starting goalies.

Your own data: If user provides CSV/JSON:

1. Verify required columns: game_date, home_team, away_team, home_goals, away_goals (to compute total), home_starter_sv_pct, away_starter_sv_pct, home_shots_per_60, away_shots_per_60, home_pp_pct, away_pp_pct, home_pk_pct, away_pk_pct
2. Verify ISO 8601 date format
3. Flag missing goalie data -- goalie features are the highest-value signal and cannot be silently imputed
4. Note: credits are not consumed

Initial Assessment

Before building, establish:

1. What is the target? (total goals, or binary over/under at a specific line like 5.5?)
2. What market is the user modeling? (DraftKings, FanDuel, Pinnacle, or multiple books?)
3. How many seasons of data are available? (totals models need 3+ seasons for stable goalie coefficients)

How It Works

Why Totals Are a Separate Model

A team can win 1-0 or win 7-5. The moneyline model doesn't care which. The totals model cares about nothing else. Different features, different calibration, different market dynamics.

Key differences from moneyline:

• Goalie quality matters more. A bad goalie matchup adds 0.5-1.0 expected goals. That moves the line by a full half-goal.
• Pace metrics replace matchup metrics. You care how fast both teams play, not which is better.
• Special teams add total goals, not change winners. PP% affects scoring rate for both teams simultaneously.
• Rest/fatigue increases scoring. Back-to-back games produce measurable total goal increases (tired defense, looser play).

Step 1: Build the Target

df['total_goals'] = df['home_goals'] + df['away_goals']

# For binary over/under at line L
df['over_5_5'] = (df['total_goals'] > 5.5).astype(int)  # over = 1
df['over_6_0'] = (df['total_goals'] > 6.0).astype(int)

Build both a regression target (total goals) and binary classification targets (over/under at each line). They inform each other during validation.

Step 2: Feature Set -- Pace Metrics

Pull get_team_stats for both teams. Apply .shift(1) before any rolling calculation.

Shots per 60 (both teams combined):

df['total_shots_per_60'] = df['home_shots_per_60_roll10'] + df['away_shots_per_60_roll10']

Higher combined shots per 60 = higher expected total. This is the primary pace driver.

Scoring chances per 60: If available, prefer scoring chances over raw shots. Scoring chances are more predictive of goals than shot counts.

Goals per game rolling average (lagged):

df['home_gf_roll10'] = home_goals shifted and rolled over 10 games
df['away_ga_roll10'] = away_goals_against shifted and rolled over 10 games
df['expected_total_simple'] = df['home_gf_roll10'] + df['away_ga_roll10']
# also compute: away_gf_roll10 + home_ga_roll10

Step 3: Feature Set -- Special Teams

PP% for both teams + PK% for both teams. These interact:

# Net PP effect on scoring rate
df['home_pp_edge'] = df['home_pp_pct_roll10'] - df['away_pk_pct_roll10']
df['away_pp_edge'] = df['away_pp_pct_roll10'] - df['home_pk_pct_roll10']
df['combined_pp_quality'] = df['home_pp_edge'] + df['away_pp_edge']

Power play opportunities per game (not just conversion rate): a team that draws more penalties gives more scoring chances regardless of PP%.

Step 4: Feature Set -- Goalie Matchup

This is the highest-leverage feature group. Goalie quality variance accounts for a larger share of total goals prediction than any other single input.

# Season SV% (lagged to current game)
df['home_starter_sv_pct'] = home_goalie_sv_pct_prior_to_game
df['away_starter_sv_pct'] = away_goalie_sv_pct_prior_to_game

# Combined goalie quality (lower = more goals expected)
df['combined_goalie_quality'] = df['home_starter_sv_pct'] + df['away_starter_sv_pct']

# Recent form (last 5 starts, by start date)
df['home_starter_sv_pct_l5'] = rolling SV% over last 5 starts for starter
df['away_starter_sv_pct_l5'] = same for away starter

# Confirmed starter flag (known before game time)
df['home_starter_confirmed'] = 1 if starter announced, 0 if uncertain

If starter is unconfirmed: Use team backup SV% as a downside scenario. Weight predictions toward higher-scoring outcomes when starter is uncertain.

Step 5: Feature Set -- Contextual Factors

Back-to-back effect:

df['home_b2b'] = (df['home_rest_days'] == 0).astype(int)
df['away_b2b'] = (df['away_rest_days'] == 0).astype(int)
df['either_b2b'] = ((df['home_b2b'] == 1) | (df['away_b2b'] == 1)).astype(int)

Back-to-back games increase total goals by approximately 0.2-0.4 goals. Tired defense and a fatigued backup goalie are the mechanisms.

Time of season:

df['games_into_season'] = game number (1-82) within the season
df['early_season'] = (df['games_into_season'] <= 20).astype(int)

Early-season games produce more goals -- goalies aren't sharp, defensive systems aren't set. This effect diminishes after ~game 20.

Head-to-head scoring history: Use get_head_to_head for last 5 meetings. High-scoring rivalry vs defensive rivalry carries predictive signal, though it's weaker than pace and goalie matchup.

Step 6: Train the Totals Model

Use XGBoost or LightGBM. Regression target preferred over binary classification -- regression gives you the full probability distribution.

from xgboost import XGBRegressor

model = XGBRegressor(
    n_estimators=400,
    max_depth=4,
    learning_rate=0.04,
    subsample=0.8,
    colsample_bytree=0.8,
    random_state=42
)
model.fit(X_train, y_train)  # y_train = total_goals

To get over/under probabilities from the regression output, assume a Poisson or normal distribution around the predicted mean:

from scipy.stats import poisson

predicted_total = model.predict(X_game)[0]
prob_over_5_5 = 1 - poisson.cdf(5, mu=predicted_total)  # P(goals >= 6)

Step 7: Calibration and the Under Bias

Documented market inefficiency: Games with total set at 5.5 go under more than the market implies. This has been documented in public NHL betting research (though its persistence varies by season and book).

Test this in your data:

line_55_games = df[abs(df['market_line'] - 5.5) < 0.1]
actual_over_rate = line_55_games['over_5_5'].mean()
implied_over_rate = 1 - line_55_games['implied_under_prob'].mean()
edge = actual_over_rate - implied_over_rate  # negative = under edge

If the pattern holds in your walk-forward validation, it's real signal. If it only shows up in-sample, it's noise.

Calibration check: Bin predicted probabilities into deciles. Check that games in the "60-70% over" bucket actually go over 60-70% of the time. Use isotonic regression or Platt scaling if calibration is off.

Step 8: Walk-Forward Validation

Use identical methodology to walk-forward-validation skill. Train on seasons 1-N, test on season N+1. Never use k-fold on time series.

For totals, the key metric is RMSE on total goals and calibrated Brier score on over/under binary targets at each line (5, 5.5, 6, 6.5).

Season Resolution

• October through December: current calendar year is the season start (2026-27 season)
• January through September: previous calendar year is the season start (2025-26 season)
• "This season" = season currently in progress or most recently completed
• NHL regular season: October to April. Playoffs: April to June.

Credit Usage

Anti-patterns

Output Format

The totals model produces:

1. Game-level predictions: predicted_total_goals (float), prob_over_5 (float), prob_over_5_5 (float), prob_over_6 (float), prob_over_6_5 (float).
2. Edge report (if odds pulled): predicted probability minus market implied probability for each line. Positive = over edge. Negative = under edge.
3. Model performance: RMSE on total goals, Brier score at each line, calibration curve per line.
4. Feature importance: ranked list showing which features drove total goals predictions in walk-forward test folds.

What to Do Next