검색 상세

AI-Based Forecasting Study : The Philadelphia Semiconductor Index （SOX）

초록/요약

Today, semiconductors are at the foundation of the modern digital economy. These essential components support artificial intelligence, cloud computing, data centers, telecommunications infrastructure, advanced manufacturing, electric vehicles, defense technologies, and the digital transformation of firms and governments. These technologies are increasingly shaping investment decisions, supply-chain strategy, and national industrial policy, making the ability to understand and predict behaviour in the semiconductor market both academically important and economically relevant. This thesis analyses the Philadelphia Semiconductor Index （SOX）, a market-based benchmark index for companies mainly engaged in the design, distribution, manufacturing, and sale of semiconductors. SOX is used as a proxy for investors' expectations regarding semiconductor demand, technological innovation, profitability, and risk at the sector level. Semiconductor equity prices are cyclical, volatile, sensitive to macroeconomic conditions, and exposed to supply-chain and geopolitical shocks, making the index difficult to forecast. In this study, we compare the following forecasting techniques: Autoregressive Integrated Moving Average （ARIMA）, Seasonal Autoregressive Integrated Moving Average （SARIMA）, Simple Recurrent Neural Network （RNN）, Long Short-Term Memory （LSTM）, Convolutional Neural Network （CNN）, and LSTCN-inspired hybrid LSTM-Conv1D （LSTCN）. The data used in this study are the daily closing prices of the SOX index from 4 January 2016 to 8 April 2026, with 2,580 observations. The data are split chronologically into an 80% training sample and a 20% testing sample, with 516 observations saved for out- of-sample evaluation. Forecast accuracy is assessed using Root Mean Square Error （RMSE）, Mean Absolute Error （MAE）, Mean Absolute Percentage Error （MAPE）, and a naive random-walk benchmark that uses the previous close to evaluate incremental predictive power. The comparative evaluation indicates that SARIMA and ARIMA are the two strongest overall forecasting models in this study. SARIMA produced the lowest RMSE, while ARIMA performed slightly better on MAE and MAPE. Among the neural network models, the returns-based RNN yielded the best result. However, its error values were very close to those of the naive random-walk benchmark, and the Diebold-Mariano test shows that its squared-error loss is not statistically different from the benchmark. This suggests that the RNN was the strongest neural network model, but it did not yield a statistically significant improvement over a simple random-walk forecast. By comparison, LSTM, the LSTCN- inspired hybrid, and CNN showed noticeably higher prediction errors, with CNN performing the worst. The results are consistent with the weak form of market efficiency but do not constitute a formal, direct test of market efficiency. Much of the historical information in SOX prices seems to be quickly adjusted into current index levels. Empirically, the study extends evidence on financial forecasting to the Philadelphia Semiconductor Index. Methodologically, it compares statistical and AI models under the same empirical design with leakage control. In practice, it shows that semiconductor forecasting systems should first be evaluated against transparent statistical and previous-close random-walk benchmarks before adopting highly advanced AI models. Keywords: Semiconductor forecasting, Philadelphia Semiconductor Index, SOX, ARIMA, SARIMA, RNN, LSTM, CNN, LSTCN, Financial Time-Series Forecasting, Artificial Intelligence, Weak-form Market Efficiency.

more

목차

1. Introduction 1
2. Literature Review 4
3. Methodology 9
3.1 Research Design and Framework 9
3.2 Forecasting Framework 10
3.3 Data Source, Study Period, and Variable Selection 12
3.4 Descriptive Data Overview 13
3.5 Data Preprocessing and Stationarity Testing 13
3.6 Train-Test Split and Leakage Control 15
3.7 Forecasting Models and Specifications 16
3.7.1 ARIMA Model 17
3.7.2 SARIMA Model 18
3.7.3 RNN Model 19
3.7.4 LSTM Model 21
3.7.5 CNN Model 22
3.7.6 LSTCN Model 24
3.7.7 Naive Random Walk 25
3.8 Forecast Evaluation and Model Comparison 26
4. Results and Discussion 27
4.1 Descriptive Movement of the SOX Series 27
4.2 Stationarity and Transformation Results 29
4.3 ARIMA Forecasting Results 31
4.4 SARIMA Forecasting Results 33
4.5 RNN Forecasting Results 35
4.6 LSTM Forecasting Results 36
4.7 CNN Forecasting Results 37
4.8 LSTCN Forecasting Results 38
4.9 Comparative Model Performance 39
4.10 Diebold-Mariano Test and Benchmark Interpretation 43
4.11 Discussion of Findings and Implications 44
5. Conclusion 45
References 49
APPENDICES 56
APPENDIX 1. Empirical Consistency and Reproducibility Note 56
APPENDIX 2. ADF Test Code 57
APPENDIX 3. ARIMA Code 65
APPENDIX 4. SARIMA Code 73
APPENDIX 5. RNN Code 89
APPENDIX 6. LSTM Code 93
APPENDIX 7. CNN Code 100
APPENDIX 8. LSTCN Code 112

more
반출 Meta View 목록