ALIs

kommt noch

Achtung: Die Navigationslinks auf dieser Seite funktionieren nicht, weil das nicht der Rahmen ist, der zu diesem Skript gehört. Wenn das Skript ordnungsgemäß dort installiert ist, wo es später laufen soll, funktionieren auch die Links in der dortigen Umgebung.

LINUX Cluster Project

Training EfficientNet and WaveNet on Chart Images & Time Series Data


Institution

  • Name: Lehrstuhl für Finanzmanagement und Kapitalmärkte
  • Address: Arcisstr. 21, 80333 München
  • Project Proposal Date: 2020-04-28 20:17:53

Abstract:

Within this project, EfficientNet B7 (https://arxiv.org/abs/1905.11946), ResNet (https://arxiv.org/abs/1512.03385) & Xception (https://arxiv.org/abs/1610.02357) will be trained for chart pattern recognition and price action prediction on a dataset comprising more than 1.2 million candlestick chart images. The images originate from stock market data of the American S&P 500, German DAX and the 3 largest cryptocurrency exchanges. In addition, WaveNet (https://arxiv.org/abs/1609.03499) and Dilated RNNs (https://arxiv.org/abs/1710.02224) will be trained on time series price data which spans over more than 100 digital asset markets for the last 7 years. The trained model will be evaluated for trend estimation and position sizing calculation as part of a time series momentum strategy for investing in digital asset markets. Since the proposed neural networks exhibit high complexity, the training process needs relatively high resources. The P100 GPU nodes of the Linux Cluster are capable of handling this load and excel in achieving shorter training times while allowing recommended capacities of the models. There may be a demand for the DGX-1 if the P100 nodes cannot cope with the capacity of the models.