TL;DR
This paper introduces a novel Fuzzy Back-Propagation Network (FBPN) model that integrates expert opinions (via Fuzzy Delphi) with Artificial Neural Networks (ANNs) to improve sales forecasting accuracy in the Printed Circuit Board (PCB) industry. The FBPN significantly outperforms traditional forecasting methods by effectively incorporating qualitative expert knowledge and handling the inherent complexities of the PCB market.
What Problem Does This Paper Address?
Accurate sales forecasting is crucial for businesses to optimize inventory, production scheduling, and overall decision-making, ultimately impacting profitability and competitiveness. The PCB industry, in particular, faces a dynamic and complex environment influenced by various market and macroeconomic factors. Traditional forecasting methods, such as time series analysis and regression, often struggle with non-linear relationships, unexpected market shifts, and the incorporation of subjective expert knowledge. While Artificial Neural Networks (ANNs) offer powerful pattern recognition capabilities, they can be data-intensive and may not effectively leverage the nuanced insights of experienced professionals. This paper aims to bridge this gap by developing a hybrid model that combines the strengths of fuzzy logic (for expert opinion integration) and ANNs (for data-driven pattern recognition) to achieve more robust and accurate sales forecasts for the PCB industry.
Methods
The core of this research is the Fuzzy Back-Propagation Network (FBPN) model, which comprises three main stages:
- Variable Selection Stage: This stage focuses on identifying the most influential input variables for forecasting. It employs two methods:
- Stepwise Regression Analysis (SRA): A statistical technique to iteratively select independent variables that best predict the dependent variable (sales).
- Fuzzy Delphi Method (FDM): A qualitative approach that incorporates expert opinions. Experts assign fuzzy numbers (linguistic terms converted to numerical ranges) to the significance of various factors. This method helps to reduce the number of questionnaires needed and avoids distorting individual expert opinions by clearly expressing the semantic structure of predicted items.
- Data Preprocessing Stage: This stage prepares the data for the FBPN.
- Rescaled Range Analysis (R/S): Used to detect the presence of trend and seasonality in the time series data.
- Winter's Exponential Smoothing: If trend and seasonality are detected, this method is applied to account for these effects before feeding the data into the network.
- FBPN Forecasting Stage: This is the core of the model where the processed data is fed into the FBPN.
- Fuzzy Input: Expert opinions, converted into fuzzy numbers, are used as inputs.
- Back-Propagation Network (BPN): The ANNs learn from the fuzzy inputs and historical data to forecast sales. The network has an input layer, a hidden layer, and an output layer.
- Defuzzification: The output signals from the network are converted back into a concrete forecasting value.
- Error Measurement: Mean Absolute Percentage Error (MAPE) is used to evaluate the forecasting accuracy.
The FBPN is designed to be very similar to a standard BPN but crucially incorporates expert opinions through fuzzy logic, allowing for a more nuanced and informed forecasting process.
Datasets
The data used in this research is derived from a real-world dataset provided by a printed circuit board company in Taiwan. The data spans from 1999/2001 to 2003/2012 and includes monthly sales amounts as the objective of the forecasting model. The input variables are collected from three domains: Market Demand, Macroeconomics, and Industrial Production.
Key Insights and Ideas
- Expert Opinion is Valuable: The research strongly emphasizes that incorporating qualitative expert knowledge through fuzzy logic significantly enhances forecasting accuracy, especially in complex industries like PCB manufacturing.
- Fuzzy Delphi is Superior for Feature Selection: The Fuzzy Delphi Method (FDM) proved to be more effective than Stepwise Regression Analysis (SRA) in selecting relevant input variables, as it can better capture the nuanced importance assigned by experts.
- Trend and Seasonality Matter: Accounting for trend and seasonality in the time series data using methods like Winter's exponential smoothing dramatically improves forecasting performance.
- Hybrid Models Outperform Traditional Ones: The proposed FBPN model, by combining fuzzy logic and ANNs, consistently outperformed traditional methods like Grey Forecasting (GF) and Multiple Regression Analysis (MRA) in terms of accuracy and error reduction.
Key Results
The FBPN model demonstrated superior performance across various evaluation metrics:
- Mean Absolute Percentage Error (MAPE): The FBPN achieved the lowest MAPE of 3.09%, significantly outperforming GF (15.04%), MRA (8.86%), and BPN (6.19%).
- Accuracy: The FBPN achieved an accuracy of 97.61%.
- Variable Selection Comparison:
- When considering trend, SRA resulted in a MAPE of 7.15%, while FDM achieved 6.19%.
- Without considering trend, SRA had a MAPE of 13.87%, and FDM had a MAPE of 12.88%.
- Encompassing Test: The FBPN and BPN models were found to encompass GF and MRA, indicating their superior ability to capture forecasting information.
Limitations and Future Work
- Limitations:
- The paper notes that the "over scale of network could prolong the executing time of training and testing and weaken the calculating ability of the network."
- While the FBPN performs well, the paper acknowledges that "gradual regression analysis may come out with a combination of various variables, which is mutually correlated."
- Future Work:
- The authors suggest that the FBPN model is a "very powerful and effective forecasting tool that can be further applied in other field of applications since expert's opinion can be incorporated into the model." This implies potential for broader application beyond the PCB industry.
- Further investigation into optimizing network architecture and learning parameters could lead to even better performance.
Summary updated