Pemodelan Harga Saham NVIDIA (NVDA) Menggunakan Model Long Short-Term Memory pada Deep Learning dengan Fitur Entropi Shannon

Abstract

Prediksi harga saham perusahaan teknologi seperti NVIDIA (NVDA) menjadi tantangan tersendiri karena karakteristik harga yang bersifat nonlinier dan volatil. Penelitian ini bertujuan mengintegrasikan Entropi Shannon sebagai fitur representasi ketidakpastian pasar ke dalam arsitektur Long Short-Term Memory (LSTM) dan membandingkan performanya dengan Random Forest (RF), LSTM baseline, dan Naïve Forecast. Data historis harga NVDA periode 2018–2024 sebanyak 1.739 observasi digunakan setelah proses rekayasa fitur yang mencakup log return, volatilitas, rolling mean, dan Entropi Shannon dengan rolling window 20 hari. Model dilatih menggunakan input sequence 30 hari serta dievaluasi menggunakan MSE, RMSE, MAE, MAPE, dan R². Hasil menunjukkan Entropi Shannon-LSTM mencapai validation loss 0,8360 pada epoch ke-8, lebih rendah dibandingkan LSTM baseline sebesar 0,8689 pada epoch ke 11. Pada data testing, model memperoleh RMSE 3,6901, MAE 2,8494, MAPE 2,5907, dan R² 0,9744, lebih unggul dibandingkan RF. Uji Diebold–Mariano menghasilkan DM 3,0636 dengan p-value 0,0022 yang signifikan pada taraf 5%.

Predicting the stock price of technology companies such as NVIDIA (NVDA) poses a challenge due to its nonlinear and volatile characteristics. This study aims to integrate Shannon Entropy as a market uncertainty feature into the Long Short-Term Memory (LSTM) architecture and to compare its performance with Random Forest (RF), baseline LSTM, and Naïve Forecast. Historical NVDA stock price data from 2018 to 2024 consisting of 1,739 observations were used after feature engineering including log return, volatility, rolling mean, and Shannon Entropy with a 20-day rolling window. The models were trained using a 30-day input sequence and evaluated using MSE, RMSE, MAE, MAPE, and R². The results show that Shannon Entropy-LSTM achieved a validation loss of 0.8360 at epoch 8, lower than the baseline LSTM of 0.8689 at epoch 11. On the testing data, the model obtained RMSE 3.6901, MAE 2.8494, MAPE 2.5907, and R² 0.9744, outperforming RF. The Diebold–Mariano test yielded DM 3.0636 with a p-value of 0.0022, significant at the 5% level.