Autoregressive Convolutional Neural Networks for Asynchronous Time Series
Authors: Mikolaj Binkowski, Gautier Marti, Philippe Donnat
ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed model on artificially generated datasets, household electric power consumption dataset available from UCI repository (Lichman, 2013), and the financial dataset of bid/ask quotes sent by several market participants active in the credit derivatives market, comparing its performance with simple CNN, singleand multilayer LSTM (Hochreiter & Schmidhuber, 1997), Phased LSTM (Neil et al., 2016) and 25-layer Res Net (He et al., 2015). |
| Researcher Affiliation | Collaboration | 1Department of Mathematics, Imperial College London, London, UK 2Hellebore Capital Limited, London, UK 3Laboratoire d informatique, Ecole Polytechnique, Palaiseau, France. |
| Pseudocode | No | The paper includes a scheme of the proposed SOCNN architecture (Figure 3) but does not present structured pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code for the experiments is available online at https://github.com/mbinkowski/nntimeseries. |
| Open Datasets | Yes | The UCI household electricity dataset7 contains measurements... available from UCI repository (Lichman, 2013) |
| Dataset Splits | Yes | The training and validation sets were sampled randomly from the first 80% of timesteps in each series, with ratio 3 to 1. The remaining 20% of data was used as a test set. |
| Hardware Specification | Yes | For artificial and electricity data we optimized the models using one K20s NVIDIA GPU while for quotes dataset we used 8-core Intel Core i7-6700 CPU machine only. |
| Software Dependencies | No | The paper states that 'Experiments were carried out using implementation relying on Tensorflow (Abadi et al., 2016) and Keras (Chollet, 2015)', but it does not specify exact version numbers for these software components. |
| Experiment Setup | Yes | We used a batch size of 128 for artificial and electricity data, and 256 for quotes dataset. We also applied batch normalization (Ioffe & Szegedy, 2015) in between each convolution and the following activation. At the beginning of each epoch, the training samples were shuffled. To prevent overfitting we applied dropout and early stopping.12 Weights were initialized following the normalized uniform procedure proposed by Glorot & Bengio (2010). For all networks we have chosen Leaky Re LU activation function (15) with leak rate a = .1 as an activation function. Table 1 presents the configurations of the network hyperparameters used in comparison (e.g., dropout rate, gradient clipping threshold). |