<![CDATA[Akurasi peramalan sebagai tolok ukur kinerja metode deret waktu bergantung beberapa hal, antara lain karakteristik data, pemilihan metode, fluktuasi data, dan keberadaaan pencilan dalam data. Keberadaan pencilan tersebut sering kali tidak dapat dihindari sehingga dapat mengganggu akurasi peramalan. Mempertimbangkan hal tersebut dalam penelitian ini dibahas tentang perbandingan kinerja metode Autoregressive Integrated Moving Average (ARIMA), Multi-Layer Perceptron (MLP), dan Random Forest (RF) dalam peramalan data deret waktu yang mengandung pencilan, menggunakan studi kasus data harga logam mulia berjangka (emas, perak, dan platina) berdasarkan nilai Mean Absolute Percentage Error (MAPE). Ditunjukkan bahwa kinerja metode ARIMA dengan Interpolasi Linier mampu menekan pengaruh pencilan lebih baik dibanding metode ARIMA dengan Winsorized Mean dan ARIMA tanpa penanganan data pencilan dengan nilai MAPE rata-rata berturut-turut sebesar 10,67% dibanding 12,33% dan 11,79% ketika dievaluasi menggunakan data uji. Metode MLP memiliki kinerja yang tidak lebih baik dibanding ARIMA dengan Interpolasi Linier dengan nilai MAPE rata-rata sebesar 11,13% ketika dievaluasi menggunakan data uji. Secara keseluruhan kinerja terbaik dihasilkan oleh metode RF, dengan nilai MAPE rata-rata jauh lebih kecil dibanding metode lainnya, yakni 2,85% ketika dievaluasi menggunakan data uji. Dalam kajian ini disimpulkan Metode RF memiliki kinerja terbaik dibandingkan semua metode. Hal tersebut disebabkan metode RF menggunakan prinsip decision tree sehingga lebih robust terhadap kehadiran pencilan dalam data. Berdasarkan hasil penelitian, metode RF dapat menjadi opsi untuk pemodelan data deret waktu yang mengandung pencilan. Abstract Forecasting accuracy as a benchmark for the performance of time series methods depends on several things, including data characteristics, method selection, data fluctuations, and the existence of outliers in the data. The existence of these outliers is often unavoidable so it can interfere with the accuracy of forecasting. Considering this, this research discusses the comparison of the performance of the Autoregressive Integrated Moving Average (ARIMA), Multi-Layer Perceptron (MLP), and Random Forest (RF) methods in forecasting time series data containing outliers, using a case study of precious metal futures price data (gold, silver, and platinum) based on the Mean Absolute Percentage Error (MAPE) value. It is shown that the performance of the ARIMA method with Linear Interpolation is able to suppress the influence of outliers better than the ARIMA method with Winsorized Mean and ARIMA without handling outlier data with the average MAPE value was obtained respectively at 10.67% compared to 12.33% and 11.79% when evaluated using test data. The MLP method has no better performance than ARIMA with Linear Interpolation with an average MAPE value of 11.13% when evaluated using test data. Overall, the best performance was produced by the RF method, which had a much smaller average MAPE value than the other methods, namely 2.85% when evaluated using test data. In this study it appears that the RF method has the best performance compared to all methods. This is because the RF method is based on decision tree principle so it is more robust to the presence of outliers in the data. Based on the research results, the RF method can be an option for modeling time series data that contains outliers. ]]>
