bikes_param_search_part2.R

results <- vector(mode = 'numeric', length = nrow(prophetGrid))

# Search best parameters
for (i in seq_len(nrow(prophetGrid))) {
    # Select the parameters to try
    parameters <- prophetGrid[i, ]
    
    if (parameters$growth == 'logistic') {
        train$cap <- parameters$capacity
    }
    
    # Initialize Prophet model
    m <- prophet(growth = parameters$growth, 
                 holidays = holidays,
                 seasonality.prior.scale = parameters$seasonality_prior_scale, 
                 changepoint.prior.scale = parameters$changepoint_prior_scale,
                 holidays.prior.scale = parameters$holidays_prior_scale,
                 yearly.seasonality = TRUE,
                 fit = FALSE)
    
    # Add exogenous features
    m <- add_regressor(m, 'weather', prior.scale = parameters$regressor_prior_scale)
    m <- add_regressor(m, 'atemp', prior.scale = parameters$regressor_prior_scale)
    m <- add_regressor(m, 'humidity', prior.scale = parameters$regressor_prior_scale)
    m <- add_regressor(m, 'windspeed', prior.scale = parameters$regressor_prior_scale)
    
    # Fit the model
    m <- fit.prophet(m, train)
    
    if (parameters$growth == 'logistic') {
        valid$cap <- parameters$capacity
    }
    
    # NOTE: There's a problem in function names with library(caret)
    forecast <- predict(m, bind_rows(train, valid))
    
    # Compute MAE and save it
    results[i] <- forecast::accuracy(forecast[ymd(forecast$ds) %in% valid$ds, ]$yhat, valid$y)[ , 'MAE']
}

prophetGrid <- cbind(prophetGrid, results)

# Choose the hyperparameters that produced the best (minimum) MAE
best_params <- prophetGrid[prophetGrid$results == min(results), ]