Date

2022-9-20

Author

TOPALLAR, SARP TUĞBERK

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

780
views

939
downloads

Cite This

Time series forecasting can be summarized as predicting the future values of a sequence indexed by timestamps based on the past records of that sequence. Optimal or near-optimal resource allocation requires accurate predictions into the future. The study presents investigation performed on both classical methods and more contemporary methods from the literature. The classical methods studied are Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS) and Seasonal-Trend decomposition using LOESS (STL). One of the more contemporary time series models is a deep learning method called DeepAR, which is a Recurrent Neural Network (RNN) comprised of Long Short Term Memory (LSTM) cells. Both novel and classical approaches pose challenges unique to their own methodologies. Classical methods for example require the data to satisfy restricting modeling assumptions and each series to be preprocessed and modeled one by one. On the other hand, machine learning methods require a great amount of quality data for training, where it would be a challenge where the abundance of data may not be available. The proposed solution is to model many similar series with a single common model at the same time. The similarity in the context of this research refers to the fact that each series has the same recorded metric. The data set investigated in the study holds records of demand for many physical stores throughout Türkiye, where each individual series corresponds to a single physical location. There are a total of 120 individual series spanning between 2018 and 2022. The probabilistic forecasts are obtained by training the DeepAR model in order to learn a probability distribution and producing the point forecasts from sampling the learned probabilistic function. The probabilistic forecasts of different quantiles provide practicality such as forecasting in different quantiles for each series individually and can be tuned for different sensitivity to forecasting errors per series. At the end of the study, the provided error metrics for accuracy measurement are MAE (Mean Absolute Error) in order to show the actual value of the forecasted demand and MAPE (Mean Absolute Percentage Error) in order to compare the results with other models independent of the scale. The DeepAR model provided more accurate results compared to ETS and ARIMA on average for the whole data set in terms of both accuracy metrics.

Subject Keywords

Time series, Demand forecasting, Deep neural networks, Long short-term memory neurons, Recurrent neural networks, Autoregressive recurrent networks

URI

https://hdl.handle.net/11511/98797

Collections

Graduate School of Applied Mathematics, Thesis