| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - bittensor |
| --- |
| ```text |
| --- |
| Copyright (c) 2023 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 |
| |
| ```text |
| 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 and dummy models in runnable/miner_testing.py. |
| ``` |
| |
| # Build Strategy |
| |
| ```text |
| This section outlines the strategy used to build the models. |
| ``` |
| |
| ## Understanding Dataset Used |
| ```text |
| The dataset used to build the models can be generated using the |
| runnable/generate_historical_data.py. A lookback period between June 2022 and |
| July 2023 on the 5m interval was used to train the model. Through analysis, the |
| reason this dataset was used is because historical data beyond June 2022 provides |
| strongly trending price movement or data movement that is from a period where |
| Bitcoin's market cap was too small to be relevant to where Bitcoin is now. |
|
|
| Therefore, using more recent data was used which correlates to the current market |
| cap and macroeconomic conditions where its uncertain we'll continue to get highly |
| trending Bitcoin data. |
|
|
| Testing data was used between June 2023 and Nov 2023 to determine performance of |
| the models. This was tested using the runnable/miner_testing.py file with a |
| separately generated test dataset from runnable/generate_historical_data.py. |
| ``` |
| |
| ## Understanding Model Creation |
| ```text |
| As of now, the TSPS infrastructure only provides close, high, low, and volume. It |
| also provides financial indicators such as RSI, MACD, and Bollinger Bands but they |
| were not used for the purposes of training these models. |
| |
| The models were derived using a variety of windows and iterations through the June |
| 2022 to June 2023 dataset. The strategy to derive the model was the following: |
| |
| base_mining_model = BaseMiningModel(len(prep_dataset.T)) \ |
| .set_neurons([[1024, 0]]) \ |
| .set_window_size(100) \ |
| .set_learning_rate(0.0000001) \ |
| .set_batch_size(500) \ |
| .set_model_dir(f'mining_models/model1.h5') |
| base_mining_model.train(prep_dataset, epochs=25) |
| |
| where an LSTM model was created by using a few or no stacked layers. Most of the |
| v4 models are actually not stacked as they performed better not being stacked for |
| the most part. This could very likely change as more feature inputs are added (this |
| is being worked on as part of the open source infra in TSPS). The window size of |
| 100 helped best predict the outcome, derived in mining_objects/base_mining_model.py |
| ``` |
| |
| ## Understanding Training Decisions |
| ```text |
| Training the model used the previous 601 rows of data as an input. This is because |
| 500 rows were used to batch, and we are looking to predict 100 rows into the future |
| (the challenge presented in the Time Series Prediction Subnet). Measures were taken |
| to ensure all data was trained on in the training data. |
| |
| Each set of 601 rows was trained on 25 times, inside another loop which iterated on |
| the entirety of the dataset from 6/22 to 6/23 50 times. This provided the model the |
| ability to get granular with details yet not overfit to any single set of rows at |
| once. Therefore, a multi-layered looping infrastructure was used to derive the models. |
| |
| for x in range(50): |
| for i in range(25): |
| train_model() |
| ``` |
| |
| ## Strategy to Predict |
| ```text |
| The strategy to predict 100 closes of data into the future was to use a 1 step |
| methodology of predicting 1 step at 100 intervals into the future and connect the |
| information by generating a line from the last close to the prediction 100 closes |
| into the future. By doing so, the model could learn to predict a single step rather |
| than all 100 where loss could continue to increase with each misstep. |
| ``` |
| |
| # Model V5 |
| ```text |
| Here's the text spaced out for readability in a README file: |
|
|
| Recommendations on how to perform better than V4 and what Model V5 will look like |
| are outlined below: |
|
|
| 1. Concentrate on more difficult moves |
| 2. Get more granular data (1m) |
| 3. Get more data sources |
| 4. Use more predicted steps |
|
|
| -- Concentrate on more difficult moves |
|
|
| The Time Series Prediction Subnet will reward models that are capable of predicting |
| more "difficult" movements in the market more than those that are less difficult. |
| Therefore, taking a strategy to train your model on larger movements or bigger |
| magnitude movements would be a good consideration. Some additional details on how |
| difficulty is calculated will be released soon but it is a combination of the |
| magnitude of the movement with the std dev of the movement in the predicted interval. |
|
|
| -- Get more granular data (1m) |
|
|
| With these larger magnitude movements, a strategy to get more granular with the data |
| would be recommended. Using 1m data to train rather than 5m would help the models |
| better predict information. |
|
|
| -- Get more data sources |
|
|
| Beyond using financial market indicators like RSI, MACD, and Bollinger Bands, the |
| TSPS open source infra will gather information for miners to help train. |
|
|
| The TSPS infrastructure will be adding data scrapers and using those data scrapers |
| to automatically gather information for you. The following pieces of information will |
| be gathered & accessible through the open source infra: |
|
|
| - Bitcoin open interest |
| - Bitcoin OHLCV data |
| - Bitcoin funding rate |
| - DXY OHLCV data |
| - Gold OHLCV data |
| - S&P 500 OHLCV data |
| - Bitcoin dominance |
| - Historical news data (sentiment analysis) |
|
|
| Using this information will provide models with information they can use to better |
| predict prices as markets correlate in movement and Bitcoin responds to other markets. |
|
|
| -- Use more predicted steps |
|
|
| Rather than only predicting a single step at the 100th predicted close in the future, |
| predict more steps. This can be achieved by training multiple models, for example, |
| 10 models each at 10 closes into the future (10, 20, 30, 40, 50, 60, 70, 80, 90, 100), |
| or by using a multi-step model with 10 steps. Both will achieve more granularity when |
| it comes to predictions and therefore can achieve a much greater RMSE score. |
|
|
| ``` |
