Regression with negative RSquared and How to Fix it

# from https://www.quora.com/When-is-R-squared-negative
set.seed(100)

df <- data.frame(x = cumsum(rnorm(100)),
                 y = cumsum(rnorm(100)))
# RJ's Addition
#df <- df[sample(1:NROW(df), size=NROW(df)), ] # Shuffler

## split between training and test ----
train_df <- df[1:80,]
test_df <- df[81:100,]

### fit OLS ----
lm.bias <- lm(y ~ x, data=train_df)
summary(lm.bias)$r.squared ## R^2 = 31.49% (in-sample)

### create a R^2 function ---
r2 <- function(data, model, response){
  preds <- predict(model, newdata=data)
  mean_response <- mean(data[,response])
  actuals <- data[,response]
  null_mse <- mean( (mean_response - actuals)^2 )  # NULL MODEL b/c mean_response is constant
  model_mse <- mean( ( preds - actuals)^2 )
  r_squared <- (null_mse - model_mse) / null_mse
  return(r_squared)
}

r2(train_df, lm.bias, "y")  ### the same as summary(); R^2 = 31.49% (in-sample) ; After reshuffling: R^2 = 33.96%

### calculate out-of-sample R^2 ----
r2(test_df, lm.bias, "y")  ## **NEGATIVE** 103% R^2 (out-sample) ; After reshuffling: R^2 = 13.74%