Enhance Model's Performance

[unibeck]

After moving to crypto the model's performance is okay (5/10). This was based on training it locally, so results are not reliable. However, now is a good time to reevaluate the model's hyper parameters and network.

• [x] 1. Download more data
  • Downloaded three+ years of data
  • Though in my testing, using more recent data is more relevant, and thus better performing, as different trading patterns have emerged (presumably). Perhaps I can identify this/create a feature from this?
• [x] 2. Use USD/BTC trading pair
  • Downloaded BTC/USDT data as it is the oldest available USD stable coin / BTC pair
• [x] 2. Run training in cloud
  • Optimized instance to use GPU
  • Solved several CPU and memory based bottlenecks
  • Updated gluonts to latest master for better support, performance, and bug fixes
• [x] 3. Tune hyperparameters
  • Implemented different levels of hyperparameters. This enables various development and production workflows
  • TODO: Run tuning in cloud
  • Tested higher number of workers for trainer. This should increase performance, but only if more than on time series is being used
• [x] 4. estimator.feat_dynamic_real
  • N/A
• [x] 5. estimator.feat_static_cat
  • N/A
• [x] 6. estimator.feat_static_real
  • Implemented other candle stick data (open, high, low)
  • TODO: Need to fix 'future values of feat_dynamic_real are assumed to be known.'
  • Shift features right and pad. This will make the model predict the next target value based on candles from a time frame's previous time frame.
• [x] 7. estimator.distr_output
  • PoissonOutput provides a much tighter probability distribution
  • LogitNormalOutput should be the preferred distribution, though it breaks with DeepAR
  • StudentT is the default and is okay
• [x] 8. Scale/normalize dataset
• [x] 9. Validation and test datasets are configured correctly
• [x] 10. Additional features. In order for the algorithm to predict movement and volatile spikes, it needs to have some sort of reasoning as to why it occurred. Try adding the following indicators:
  • ATR
  • Bollinger Bands
  • Momentum indicators such as ADX?
  • etc...
• [x] 11. DeepAR is useful for seasonality, though a better model would probable be LSTM See https://aws.amazon.com/blogs/machine-learning/training-debugging-and-running-time-series-forecasting-models-with-the-gluonts-toolkit-on-amazon-sagemaker/
• [x] 12. Using different timeframes as covariates. For example, use 1 and 5 minutes time series as covariates as models will learns the embeddings/relationships between these two

[unibeck]

Determine if we are using the full dataset. For example, if the parameters are epochs=100, batch_size=128, num_batches_per_epoch=50, then that is 100 x 50 x 128 = 640,000 data points used. We currently have 265,486 data points in the train dataset.