The client was a bank who wanted to specifically use deep learning methods on high frequency bond data. DataSpartan brought in a consultant with a mathematics PhD to deliver on the requirements.
Multiple research papers were sifted through and organised – the consultant liaised with various university professors to narrow down the scope of papers and consulted internally with client staff to determine the optimal method. Deep recurrent neural networks were chosen and applied to analyse the data. The Python library Tensorflow and the Spark framework was used alongside the NVIDIA Cuda Library to code the back tester.
The trading system has been put into production and a total of £3 million in capital has been allocated for the strategy. Further tweaks and improvements on the variables are being done as well to improve historic performance.