Time series analysis : forecasting and control.Įxponential Smoothing for Predicting Demand. M⁴ Competition – The M⁴ (Makridakis 4) Forecasting Competition. We anticipate that our contribution will enable a strong adoption of state-of-the-art algorithms over univariate series, and facilitate the generalization of the model towards specific problems where covariates are available. We obtained similar results to those of the original submission, but with orders of magnitude less training time. Our work focused on the vectorization of the per-time series parameters to enable GPU computation in a framework that supports eager execution (e.g. This proved difficult due to the direct implementation of the original submission on a CPU. The main challenge we overcame was the training of per-time series parameters. In this project, we successfully implement the state-of-the-art ES-RNN algorithm in a fast, accessible, and generalizable forecasting framework. Table 6: Breakdown of sMAPE by time period and category 7 Discussion Table 5: Comparison of run-times for 15 epochs Data Category Table 4: Comparison of results to the M4 baseline model Time Period Finally, Table 6 breaks down the sMAPE achieved by our implementation in each of the frequencies (and corresponding models) for the six different data categories.
TIMENET TIME SERIES CLASSIFICATION DIFFERENTIABLE UPDATE
( 2018) at this time and we will update the table when we have their reports. 3 3 3We were not able to get the running time on yearly data from Smyl et al. ( 2018) were running the algorithm of 6 pairs of 2 workers and for quarterly data, 4 pairs of 2 workers were used. Furthermore, Table 5 shows the running times of our implementation on the GPU and the running time of Smyl et al. ( 2018), that of Hyndman ( 2018) and our implementation of Smyl et al. Table 4 shows the results of the ES-RNN model by Smyl et al. Finally, we have a simple linear layer at the end for adapting the RNN output to the output prediction window in the form of normalized and de-seasonalized data, an important distinction which we discuss further in subsection 3.4. The RNN layers also have a residual connections that helps stabilize the network’s training. This means that the first LSTM hidden weights are inputs to the next cell in the layer, however the second layer has a dilation of two which means that the hidden weights and the bias weights are forwarded two cells forward and so on and so forth. For example, in Figure 1, the RNN has dilations (1, 2) and (4, 8). The main advantages of this LSTM structure, when compared to a vanilla LSTM, are that it greatly increases computational efficiency and allows the network to remember information from earlier time instances. The model at its base consists of a LSTM layer with skip connections to form the Dilated LSTM network as described in Chang et al. A summary of the different models for quarterly, monthly and yearly time periods is presented in Table 1. The neural network architectures used by Smyl et al.
0 kommentar(er)
