This paper by Kulkarni and Haidar explores the use of artificial neural networks (ANNs) to forecast short-term (up to three days) crude oil spot price direction, examining whether incorporating crude oil futures prices improves forecasting accuracy. The central question revolves around whether futures prices contain new information that can enhance spot price prediction.
Problem Statement and Motivation:
The authors highlight the economic significance of crude oil and the importance of forecasting its price direction to mitigate the negative impacts of price fluctuations. They note the difficulty in using fundamental variables for daily forecasting due to data availability. Given the dynamic nature of the oil market, influenced by political events, weather, and speculation, accurate short-term price prediction is crucial for investors and risk management.
Methodology:
The authors adopt a three-layer feedforward neural network with backpropagation. Their approach involves:
- Benchmark Creation: Establishing a baseline model using lagged values of pre-processed spot prices. The research determines an optimal number of lags, based on the historical spot price data alone.
- Futures Price Integration: Systematically adding pre-processed futures prices (1, 2, 3, and 4 months to maturity) to the benchmark model, both individually and collectively, to assess their impact on forecast accuracy.
- Data Pre-processing: Employs several pre-processing techniques to transform the raw data into stationary form. These included first and second-order relative changes. The paper finds that the combination of using "momentum" and "force" equations derived from relative changes produces better results. In addition, a 3-day moving average filter is applied to reduce noise.
- Model Optimization: Dedicated effort is spent on finding the optimal ANN model structure, including the number of hidden neurons, activation functions, learning rate, and optimization algorithm. Levenberg-Marquardt optimization algorithm was selected due to its fast convergence.
Key Design Considerations for the ANN:
- Convergence: Ensuring the model accurately fits the training data.
- Generalization: Evaluating the model's ability to perform well on unseen data.
- Stability: Verifying the consistency of the network's output across multiple runs.
Performance Metrics:
The primary metric for evaluating the model's performance is the "hit rate," or the success ratio for correctly predicting the direction of the price change. Secondary metrics include Root Mean Squared Error (RMSE), correlation coefficient (R and R<sup\>2</sup>), Mean Squared Error (MSE), Mean Absolute Error (MAE), Sum Squared Error (SSE), and the Information Coefficient (IC).
Data:
The paper uses daily closing prices for West Texas Intermediate (WTI) crude oil spot prices and futures contracts traded on the NYMEX from September 1996 to August 2007 (2705 data points). 90% of the data is used for training, and 10% for out-of-sample testing.
Results and Discussion:
- The benchmark model, based on lagged spot prices, achieved a direction forecast accuracy of around 81.38% in-sample with 20 lags, and 77.55% out-of-sample with 3-day moving average filter and data transformation, suggesting that market dynamics are best represented by more lags. The research suggested 13 lags are optimal to forecast spot price direction for the short-term.
- The results on integrating future data, shows that on the short-term, futures prices do hold new information on the spot price direction. Futures contracts 1, 2 have performed better than contracts 3, 4 but the overall improvement was insignificant.
- For all the experiments, that include futures data as an input, the results show that on the short-term, futures prices do hold new information on the spot price direction.
- Multi-step forecasts (up to three days ahead) revealed acceptable accuracy for one and two-day predictions.
Conclusion:
The paper concludes that ANN models can be used to forecast crude oil spot price direction in the short term. It provided evidence that crude oil futures prices contain new information regarding spot price direction, but the improvement achieved by adding futures data to the benchmark was relatively modest. The paper suggests that the relationship between spot and futures prices could be different during the day, warranting future research with intraday data (if available). The authors also propose investigating other variables, such as heating oil prices, interest rates, and gold prices, to further enhance short-term forecasting accuracy.
Literature Review Highlights:
The authors also provide a brief literature review, highlighting the existing research on crude oil price forecasting, including studies using ARMA, GARCH, SVM, and hybrid models. They point out the limitations of some studies, such as using raw data without pre-processing or employing old data for training ANNs. The authors also acknowledge the inconsistency in findings regarding the relationship between spot and futures prices in the existing literature.