Predicting MLB Wins: A Multiple Linear Regression Analysis
Sheriann McLarty
September 29, 2025
Executive Summary
This report develops regression models to predict MLB team wins using
batting, pitching, and fielding statistics (1871–2006).
The strongest model, which balances batting, pitching, and fielding
variables, explains 24% of the variation in wins.
Key Insights: - More Hits and Walks
drive wins.
- More Errors cost wins.
- Efficiency metrics (e.g., HR rate) confirm the Moneyball philosophy
that smarter play matters more than raw totals.
Recommendation: Use the balanced model for predictions, while exploring advanced metrics in future work.
1. Data Exploration
## [1] 2276 17## [1] 259 16## Min. 1st Qu. Median Mean 3rd Qu. Max.
## 0.00 71.00 82.00 80.79 92.00 146.00
| variable | pct_missing |
|---|---|
| TEAM_BATTING_HBP | 91.6% |
| TEAM_BASERUN_CS | 33.9% |
| TEAM_FIELDING_DP | 12.6% |
| TEAM_BASERUN_SB | 5.8% |
| TEAM_BATTING_SO | 4.5% |
| TEAM_PITCHING_SO | 4.5% |
| INDEX | 0.0% |
| TARGET_WINS | 0.0% |
We focus on Hits and Walks because correlation analysis shows
they’re strongly related to Wins.
2. Data Preparation
| Variable | Mean | Median | SD |
|---|---|---|---|
| TARGET_WINS | 80.79 | 82.00 | 15.75 |
| TEAM_BATTING_H | 1469.27 | 1454.00 | 144.59 |
| TEAM_BATTING_BB | 501.56 | 512.00 | 122.67 |
| TEAM_FIELDING_E | 246.48 | 159.00 | 227.77 |
| TEAM_PITCHING_SO | 817.54 | 813.50 | 540.54 |
| HR_rate | 0.07 | 0.07 | 0.04 |
| BB_rate | 0.34 | 0.36 | 0.09 |
| SO_rate | 0.51 | 0.53 | 0.18 |
| Net_Steals | 71.88 | 52.00 | 83.06 |
| Net_BB | -51.45 | -24.00 | 150.83 |
Median imputation was chosen because it’s robust to outliers in count data.
3. Build Models
##
## Regression Models
## ================================================================================================
## Dependent variable:
## ----------------------------------------------------------------------------
## TARGET_WINS
## (1) (2) (3)
## ------------------------------------------------------------------------------------------------
## TEAM_BATTING_H 0.048*** 0.051***
## (0.003) (0.003)
##
## TEAM_BATTING_HR 0.001 -0.026
## (0.008) (0.024)
##
## TEAM_BATTING_BB 0.030*** 0.017***
## (0.003) (0.003)
##
## TEAM_BATTING_SO 0.005** 0.001
## (0.002) (0.002)
##
## TEAM_FIELDING_E -0.016***
## (0.002)
##
## HR_rate 109.298***
## (12.069)
##
## BB_rate -3.706
## (4.751)
##
## SO_rate -45.808***
## (2.857)
##
## Net_Steals 0.061***
## (0.004)
##
## Net_BB 0.006*
## (0.003)
##
## TEAM_PITCHING_SO 0.001* 0.003***
## (0.001) (0.001)
##
## TEAM_PITCHING_HR 0.015
## (0.021)
##
## log_E -10.056***
## (0.978)
##
## Constant -8.918* 0.557 144.152***
## (4.717) (4.866) (6.486)
##
## ------------------------------------------------------------------------------------------------
## Observations 2,276 2,276 2,276
## R2 0.224 0.246 0.187
## Adjusted R2 0.223 0.243 0.185
## Residual Std. Error 13.887 (df = 2271) 13.702 (df = 2268) 14.223 (df = 2268)
## F Statistic 164.078*** (df = 4; 2271) 105.547*** (df = 7; 2268) 74.666*** (df = 7; 2268)
## ================================================================================================
## Note: *p<0.1; **p<0.05; ***p<0.01Model 1: simple batting baseline.
Model 2: strongest performance, interpretable.
Model 3: efficiency-focused, but slightly weaker fit
4. Select Model
| Model | Adj_R2 | AIC | RMSE |
|---|---|---|---|
| M1 | 0.223 | 18441.97 | 13.871 |
| M2 | 0.243 | 18383.96 | 13.678 |
| M3 | 0.185 | 18553.77 | 14.198 |
Decision: Model 2 is selected — best Adj. R², lowest
RMSE, and interpretable:
- Hits (+), Walks (+), Errors (−), Pitcher SO (+).
- HR coefficient negative due to multicollinearity, but overall model
fit is best.
Predictions
| INDEX | PREDICTED_WINS |
|---|---|
| 9 | 68.89 |
| 10 | 70.43 |
| 14 | 78.62 |
| 47 | 85.61 |
| 60 | 67.77 |
| 63 | 69.31 |
Key Takeaways & Future Work
Takeaways - Teams win more games when they
hit more and walk more.
- Errors directly cost wins and are a controllable
weakness.
- Efficiency-based features validate the Moneyball thesis, but raw
batting + pitching still outperform them in predictive power.
Future Work - Apply regularization
(ridge/lasso) to reduce multicollinearity.
- Add interaction terms (e.g., HR × SO).
- Integrate modern sabermetric stats (OPS, WAR) for
stronger prediction.
Appendix — Full Code
Appendix A: R Code
## Appendix A: Full R Code
# =========================
# Libraries
# =========================
library(tidyverse); library(corrplot); library(caret); library(car)
library(stargazer); library(kableExtra); library(broom)
# =========================
# Data Load
# =========================
train <- read.csv("moneyball-training-data.csv")
eval <- read.csv("moneyball-evaluation-data.csv")
dim(train); dim(eval)
summary(train$TARGET_WINS)
# =========================
# Missingness
# =========================
pct_na <- function(x) mean(is.na(x))*100
missing_tbl <- train %>%
summarise(across(everything(), pct_na)) %>%
pivot_longer(everything(), names_to="variable", values_to="pct_missing") %>%
arrange(desc(pct_missing))
# =========================
# Correlation Matrix
# =========================
num_vars <- train %>% select(-INDEX)
cor_mat <- cor(num_vars, use="pairwise.complete.obs")
corrplot(cor_mat, method="color", tl.cex=0.7)
# =========================
# Imputation & Feature Engineering
# =========================
preproc <- preProcess(train, method=c("medianImpute"))
train_imp <- predict(preproc, train)
eval_imp <- predict(preproc, eval)
safe_div <- function(num, den) ifelse(den==0|is.na(den), 0, num/den)
featurize <- function(df){
df %>%
mutate(
HR_rate = safe_div(TEAM_BATTING_HR, TEAM_BATTING_H),
BB_rate = safe_div(TEAM_BATTING_BB, TEAM_BATTING_H),
SO_rate = safe_div(TEAM_BATTING_SO, TEAM_BATTING_H),
Net_Steals = TEAM_BASERUN_SB - TEAM_BASERUN_CS,
Net_BB = TEAM_BATTING_BB - TEAM_PITCHING_BB,
log_E = log1p(TEAM_FIELDING_E)
)
}
train_feat <- featurize(train_imp)
eval_feat <- featurize(eval_imp)
# =========================
# Descriptive Statistics
# =========================
desc_tbl <- train_feat %>%
select(TARGET_WINS, TEAM_BATTING_H, TEAM_BATTING_BB, TEAM_FIELDING_E,
TEAM_PITCHING_SO, HR_rate, BB_rate, SO_rate, Net_Steals, Net_BB) %>%
summarise(across(everything(),
list(Mean=mean, Median=median, SD=sd),
.names="{.col}_{.fn}")) %>%
pivot_longer(everything(),
names_to=c("Variable",".value"),
names_pattern="(.*)_(Mean|Median|SD)")
# =========================
# Model Building
# =========================
m1 <- lm(TARGET_WINS ~ TEAM_BATTING_H + TEAM_BATTING_HR +
TEAM_BATTING_BB + TEAM_BATTING_SO,
data=train_feat)
m2 <- lm(TARGET_WINS ~ TEAM_BATTING_H + TEAM_BATTING_HR +
TEAM_BATTING_BB + TEAM_BATTING_SO +
TEAM_FIELDING_E + TEAM_PITCHING_SO + TEAM_PITCHING_HR,
data=train_feat)
m3 <- lm(TARGET_WINS ~ HR_rate + BB_rate + SO_rate +
Net_Steals + Net_BB + TEAM_PITCHING_SO + log_E,
data=train_feat)
stargazer(m1, m2, m3, type="text")
# =========================
# Model Comparison
# =========================
cmp <- tibble(
Model=c("M1","M2","M3"),
Adj_R2=c(summary(m1)$adj.r.squared,
summary(m2)$adj.r.squared,
summary(m3)$adj.r.squared),
AIC=c(AIC(m1),AIC(m2),AIC(m3)),
RMSE=c(RMSE(predict(m1,train_feat),train_feat$TARGET_WINS),
RMSE(predict(m2,train_feat),train_feat$TARGET_WINS),
RMSE(predict(m3,train_feat),train_feat$TARGET_WINS))
)
# =========================
# Diagnostics
# =========================
par(mfrow=c(2,2)); plot(m2); par(mfrow=c(1,1))
# =========================
# Predictions
# =========================
pred_out <- eval %>%
mutate(PREDICTED_WINS = predict(m2, newdata=eval_feat)) %>%
select(INDEX,PREDICTED_WINS)
write.csv(pred_out,"moneyball_predictions.csv",row.names=FALSE)