|
--- |
|
license: mit |
|
language: |
|
- en |
|
tags: |
|
- bittensor |
|
--- |
|
MIT License |
|
|
|
Copyright (c) 2024 Taoshi Inc |
|
|
|
Permission is hereby granted, free of charge, to any person obtaining a copy |
|
of this software and associated documentation files (the "Software"), to deal |
|
in the Software without restriction, including without limitation the rights |
|
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
|
copies of the Software, and to permit persons to whom the Software is |
|
furnished to do so, subject to the following conditions: |
|
|
|
The above copyright notice and this permission notice shall be included in all |
|
copies or substantial portions of the Software. |
|
|
|
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
|
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
|
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
|
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
|
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
|
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
|
SOFTWARE. |
|
|
|
# Background |
|
|
|
The models provided here were created using open source modeling techniques |
|
provided in https://github.com/taoshidev/time-series-prediction-subnet (TSPS). |
|
They were achieved using the `runnable/miner_training.py`, and tested against |
|
existing models in `runnable/miner_testing.py`. |
|
|
|
> **Note**<br> |
|
This model requires the Feature Set Creator (FSC) functionality added in the |
|
latest release of the TSPS. |
|
|
|
# Build Strategy |
|
|
|
This section outlines the strategy used to build the models. |
|
|
|
## Understanding Dataset Used |
|
|
|
The dataset used to build the models can be generated using the |
|
`runnable/generate_historical_data.py`. A lookback period between June 2023 and |
|
January 2024 on the 5m interval was used to train the model. Recent data was |
|
used because it more closely correlates to the current market and |
|
macroeconomic conditions. |
|
|
|
Testing data was used between January 2024 and February 2024 to determine the |
|
performance of the models. This was tested using the `runnable/miner_testing.py` |
|
file with live historical data sources. |
|
|
|
|
|
## Understanding Model Creation |
|
|
|
As of now, the model only uses the following features to predict: |
|
- close |
|
- high |
|
- low |
|
- volume |
|
- time of day |
|
- time of week |
|
- time of month |
|
|
|
Other features from a wide range of feature sources are being added to TSPS |
|
infrastructure in the near future as improvements to the FSC. |
|
|
|
A variety of windows and parameters were tested and eliminated. The final |
|
strategy to derive this model was the following: |
|
|
|
``` |
|
model = BaseMiningModel( |
|
filename="model_v5_1.h5", |
|
mode="w", |
|
feature_count=7, |
|
sample_count=500, |
|
prediction_feature_count=1, |
|
prediction_count=10, |
|
prediction_length=100, |
|
layers=[ |
|
[1024, 0], |
|
[1024, 0.3], |
|
], |
|
learning_rate=0.000001, |
|
dtype=Policy("mixed_float16"), |
|
) |
|
``` |
|
|
|
The LSTM model has two stacked layers with a 0.3 dropout rate. |
|
|
|
## Understanding Training Decisions |
|
|
|
Training was done with 500 samples per scenario and 128 scenarios per batch, |
|
with 20 training epochs and 10 passes over the entire dataset. Additional |
|
epochs and passes were not found to improve the model's predictions. |
|
|
|
## Strategy to Predict |
|
|
|
The strategy to predict 100 closes of data into the future was to use 10 |
|
predictions evenly spaced along the length of the prediction space, and then |
|
linearly interpolating between each prediction. By doing so, the model could |
|
learn to predict the general shape of the market movement, rather than |
|
predicting all 100. |
|
|
