Energy Price Forecasting Model

econometrics
forecasting
machine-learning
tutorial
Energy price forecasting model using AutoARIMA and model training for class example.
Author

Iván de Luna-Aldape

Published

July 23, 2024

knitr::opts_chunk$set(echo= TRUE)

Load required libraries

library(quantmod)
library(forecast)
library(ggplot2)
library(lubridate)

Download data

In this section, we’ll download the historical price data for crude oil futures:

download_data <- function(symbol, start_date, end_date){
  getSymbols(symbol, from = start_date, to = end_date, auto.assign = FALSE)
}

# Set parameters
symbol = "CL=F"
start_date = "2020-01-01"
end_date = "2023-12-31"

# Download data
data <- download_data(symbol, start_date, end_date)
Warning: CL=F contains missing values. Some functions will not work if objects
contain missing values in the middle of the series. Consider using na.omit(),
na.approx(), na.fill(), etc to remove or replace them.

Downladed data containd an index, which is in date format but must be adapted to time series, also contains prices in open, high, low and close data, volume which informs about the quantity traded and adjusted close prices.

cat("Dataset summary:\n")
Dataset summary:
print(summary(data))
     Index              CL=F.Open        CL=F.High         CL=F.Low     
 Min.   :2020-01-02   Min.   :-14.00   Min.   : 13.69   Min.   :-40.32  
 1st Qu.:2020-12-30   1st Qu.: 53.35   1st Qu.: 53.84   1st Qu.: 52.61  
 Median :2021-12-30   Median : 72.97   Median : 73.95   Median : 71.65  
 Mean   :2021-12-31   Mean   : 69.83   Mean   : 71.22   Mean   : 68.30  
 3rd Qu.:2022-12-29   3rd Qu.: 83.22   3rd Qu.: 84.55   3rd Qu.: 81.82  
 Max.   :2023-12-29   Max.   :124.66   Max.   :130.50   Max.   :120.79  
                      NA's   :1        NA's   :1        NA's   :1       
   CL=F.Close      CL=F.Volume      CL=F.Adjusted   
 Min.   :-37.63   Min.   :      0   Min.   :-37.63  
 1st Qu.: 53.34   1st Qu.: 287602   1st Qu.: 53.34  
 Median : 72.87   Median : 352450   Median : 72.87  
 Mean   : 69.77   Mean   : 378186   Mean   : 69.77  
 3rd Qu.: 83.13   3rd Qu.: 436951   3rd Qu.: 83.13  
 Max.   :123.70   Max.   :2288230   Max.   :123.70  
 NA's   :1        NA's   :1         NA's   :1       

We can also check if the time period that was downloaded is the one we specified

cat("\nTime period covered:\n")

Time period covered:
cat("Start date:", as.Date(index(data)[1]), "\n")
Start date: 18263 
cat("End date:", as.Date(index(data)[nrow(data)]),"\n")
End date: 19720 
cat("Total number of trading days:", nrow(data),"\n")
Total number of trading days: 1007 

Preprocess data

Now, let’s process the data by extracting the closing prices and ensuring we have a continous daily time series:

preprocess_data <- function(data){
  # Extract closing prices and convert to time series
  close_prices <- as.numeric(data[,4])
  dates <- index(data)
  ts_data <- xts(close_prices, order.by = dates)
  
  # Ensure daily frequency and forward fill missing values
  daily_data <- merge(ts_data, xts(, seq(start(ts_data), 
                                         end(ts_data), 
                                         by="day")))
  daily_data <- na.locf(daily_data)
  
  return(daily_data)
}

processed_data <- preprocess_data(data)

We can take a look at our data by ploting the time series:

ggplot(data.frame(Date = index(processed_data), 
                  Price = as.numeric(processed_data)), 
       aes(x = Date, y = Price)) +
  geom_line() + 
  labs(title = "Crude oil Futures Closing Prices",
       x = "Date",
       y = "Price (USD)") +
  theme_minimal()

Calculate and display some statistics

cat("\nBasic statistics of closing prices:\n")

Basic statistics of closing prices:
print(summary(as.numeric(processed_data)))
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 -37.63   53.34   72.91   69.86   83.19  123.70 

The graph shows the daily closing prices of crude oil futures over the specified time period. We can observe serevail intersting details:

  1. A sharp drop in prices in early 2020, likely due to the COVID-19 pandeminc.
  2. A gradual recovery and upward trend from mid-2020 to mid-2022.
  3. Some volatility and a slight downward trend in the later part of 2022 and 2023.

These patterns and the overall volatility will be challenging for our forecasting model to capture.

Split data

We’ll now split our data into training and testing sets:

split_data <- function(data, test_size = 0.2){
  split_point <- floor(nrow(data) * (1 - test_size))
  train <- data[1:split_point, ]
  test <- data[(split_point + 1):nrow(data), ]
  return(list(train = train, test = test))
}

split <- split_data(processed_data)
train <- split$train
test <- split$test

cat("Training set size:", nrow(train), "days\n")
Training set size: 1166 days
cat("Test set size:", nrow(test), "days\n")
Test set size: 292 days

We’ve split the data so that 80% is used for training and 20% for testing. This allows us to train our model on a substantial amount of historical data while still having a significant portion for evaluating its performance.

Train SARIMA Model

Now, let’s train a SARIMA (Seasonal AutoRegressive Integrated Moving Average) model:

train_sarima <- function(data){
  # Automatically select the best SARIMA model
  model <- auto.arima(data)
  return(model)
}

model <- train_sarima(train)

summary(model)
Series: data 
ARIMA(0,1,1) 

Coefficients:
          ma1
      -0.3158
s.e.   0.0293

sigma^2 = 7.066:  log likelihood = -2791.57
AIC=5587.14   AICc=5587.15   BIC=5597.26

Training set error measures:
                     ME   RMSE      MAE       MPE     MAPE     MASE       ACF1
Training set 0.01931691 2.6559 1.228565 0.4895984 2.340853 1.131177 0.01242817

SARIMA is a popular model for time series forecasting, with the following properties:

  1. Seasonal component: it can capture both seasonal and non-seasonal patterns in the data.
  2. Autoregressive (AR): It uses past values to predict future values.
  3. Integrated (I): It can make the time series stationary by differencing.
  4. Moving Average (MA): It uses past forecast erros in the prediction equation.

The auto.arima() function automatically selects the best SARIMA model based on the AIC (Akaike Information Criterion). The model summary shows:

  • The selected ARIMA order (p,d,q): (0, 1, 1)
  • Coefficients in the model: which is an ma1 of -0.3158
  • Measures of fit like AIC, BIC, and log-likelihood

This automated approach saves time in model selection, but it’s always good to validate the results and potentially try manual parameter tuning if needed.

Make predictions

Let’s use our trained model to make predictions:

make_predictions <- function(model, n_periods){
  forecast(model,  h = n_periods)
}

predictions <- make_predictions(model, nrow(test))

The forecast() function generates pint forecasts as well as prediction intervals. The prediction intervals give us an idea of the uncertainty in our forecasts.

cat("Predictions mean value\n:")
Predictions mean value
:
head(predictions$mean)
Time Series:
Start = 1167 
End = 1172 
Frequency = 1 
[1] 76.6646 76.6646 76.6646 76.6646 76.6646 76.6646
cat("Predictions lower values\n:")
Predictions lower values
:
head(predictions$lower)
Time Series:
Start = 1167 
End = 1172 
Frequency = 1 
          80%      95%
1167 73.25800 71.45466
1168 72.53698 70.35195
1169 71.92439 69.41508
1170 71.38237 68.58613
1171 70.89102 67.83467
1172 70.43832 67.14232
cat("Predictions upper values\n:")
Predictions upper values
:
head(predictions$upper)
Time Series:
Start = 1167 
End = 1172 
Frequency = 1 
          80%      95%
1167 80.07119 81.87454
1168 80.79222 82.97725
1169 81.40481 83.91412
1170 81.94682 84.74307
1171 82.43818 85.49453
1172 82.89088 86.18688

Evaluate Model

Now, let’s evaluate our model’s performance:

evaluate_model <- function(actual, predicted){
  rmse <- sqrt(mean((actual - predicted$mean)^2))
  mae <- mean(abs(actual - predicted$mean))
  mape <- mean(abs((actual - predicted$mean) / actual)) * 100
  return(list(RMSE = rmse, MAE = mae, MAPE = mape))
}

metrics <- evaluate_model(test, predictions)
cat("RMSE:", metrics$RMSE, "\n")
RMSE: 6.643713 
cat("MAE:", metrics$MAE, "\n")
MAE: 5.556715 
cat("MAPE:", metrics$MAPE, "\n")
MAPE: 7.044422 

We are using three common metrics to evaluate our model:

  1. Root Mean Square Error (RMSE): Measures the standard deviation of the residuals.
  2. Mean Absolute Error (MAE): Measures the average magnitude of the errors in a set of predictions.
  3. Mean Absolute Percentage Error (MAPE): Measures accuracy as a percentage, giving us an idea of how far the predictions are off on average.

These metrics give us different perspectives on our model’s performance. Lower values indicate better performance.

Plot results

Finally, let’s visualize our results:

plot_results <- function(train, test, forecast){
  # Combine data
  all_data <- rbind(train, test)
  
  # Create a data frame for ggplot
  plot_data <- data.frame(
    Date = index(all_data),
    Price = as.numeric(all_data),
    Type = c(rep("Train", nrow(train)), rep("Test", nrow(test)))
  )
  
  forecast_data <- data.frame(
    Date = index(test),
    Price = as.numeric(forecast$mean),
    Type = "Forecast"
  )
  
  plot_data <- rbind(plot_data, forecast_data)
  
  # Create the plot
  ggplot(plot_data, aes(x = Date, y = Price, color = Type)) + 
    geom_line() +
    geom_ribbon(data = forecast_data,
                aes(ymin = forecast$lower[,"95%"],
                    ymax = forecast$upper[,"95%"],
                    fill = Type),
                alpha = 0.2) +
    labs(title = "Energy Price Forecasting",
         x = "Date",
         y = "Price") +
    theme_minimal()
}

plot_results(train, test, predictions)

This plot shows:

  • The original training data in blue
  • The actual test data in green
  • Our model’s predictions in green
  • The shaded area represents the 95% prediction interval for our forecasts