| --- |
| 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. |
|
|
