Electronic Theses and Dissertation

Harga saham perusahaan platform streaming global berfluktuasi tinggi akibat
dinamika pasar yang stokastik dan pengaruh sentimen investor, sehingga prediksi harga
penutupan menjadi penting untuk mendukung pengambilan keputusan investasi.
Penelitian ini bertujuan membandingkan performa Long Short-Term Memory (LSTM),
Gated Recurrent Unit (GRU), dan Transformer dalam memprediksi harga penutupan
saham Netflix (NFLX), The Walt Disney Company (DIS), Warner Bros. Discovery
(WBD), Roku, Inc. (ROKU), dan Amazon.com, Inc. (AMZN) pada periode 28
September 2017 hingga 13 Maret 2026. Penelitian menerapkan skema direct multi-step
forecasting dengan input window 60, 90, dan 120 hari serta output 5, 10, 20, 30, dan 45
hari ke depan. Model terbaik dari setiap kombinasi dioptimasi menggunakan Optuna
dengan algoritma Tree-structured Parzen Estimator (TPE) sebanyak 20 trial. Evaluasi
dilakukan menggunakan metrik Root Mean Squared Error (RMSE), Mean Absolute
Error (MAE), Mean Absolute Percentage Error (MAPE), training time, dan inference
time. Hasil menunjukkan bahwa performa model bersifat kontekstual dan bergantung
pada karakteristik saham serta horizon prediksi. Pada saham NFLX dan AMZN, GRU
pasca-tuning cenderung menghasilkan performa terbaik. Pada saham DIS, WBD, dan
ROKU, model terbaik bervariasi menurut horizon prediksi, sehingga tidak terdapat satu
arsitektur yang unggul secara universal. Secara umum, Transformer menghasilkan nilai
MAPE yang lebih tinggi dibandingkan LSTM dan GRU pada sebagian besar
konfigurasi. Hasil visualisasi juga menunjukkan bahwa nilai MAPE cenderung
meningkat seiring bertambahnya horizon prediksi, mencerminkan akumulasi
ketidakpastian pada skema direct multi-step forecasting.
Kata kunci: harga saham, platform streaming global, direct multi-step forecasting,
LSTM, GRU, Transformer

Global streaming platform companies stock prices fluctuate due to stochastic market dynamics and investor sentiment, making closing price prediction important for investment decision-making. This study compares the performance of Long Short- Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer in predicting the closing stock prices of Netflix (NFLX), The Walt Disney Company (DIS), Warner Bros. Discovery (WBD), Roku, Inc. (ROKU), and Amazon.com, Inc. (AMZN) over the period from September 28, 2017 to March 13, 2026. The study applies a direct multi-step forecasting scheme with input windows of 60, 90, and 120 days and outputs of 5, 10, 20, 30, and 45 days ahead. The best model from each combination was optimized using Optuna with the Tree-structured Parzen Estimator (TPE) algorithm for 20 trials. Evaluation was conducted using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), training time, and inference time. The results indicate that model performance is contextual and depends on stock characteristics and prediction horizons. For NFLX and AMZN, post-tuning GRU tended to achieve the best performance. For DIS, WBD, and ROKU, the best model varied across prediction horizons, indicating that no single architecture was universally superior. In general, Transformer produced higher MAPE values than LSTM and GRU across most configurations. The visualization results also show that MAPE tends to increase as the prediction horizon lengthens, reflecting the accumulation of uncertainty in the direct multi-step forecasting scheme. Keywords: stock price, global streaming platforms, direct multi-step forecasting, LSTM, GRU, Transformer