This project uses two neural networks to predict Bitcoin's direction of movement (up, down, flat) based on technical indicators calculated using OHLCV data from Bybit (spot trading, 30-minute candles). Technologies: Python, TensorFlow/Keras, NumPy, Pandas, Matplotlib, Scikit-learn,Technical Analysis (ta), requests, keras_tuner
Source: Bybit, spot trading
Format: OHLCV (Open, High, Low, Close, Volume) candles, 30-minute
Raw data is fetched and processed using get_raw_candles_csv in data\raw_data\get_raw_candle.py from Bybit API.
Technical indicators:
The dataset is prepared using processing_data_c5 in file:data\ready_data\processing_raw_candle.py, which calculates technical indicators and creates a training CSV file.
| Indicator | Description |
|---|---|
| Close_t, Close_(t-1), Close_(t-2), Close_(t-3), Close_(t-4) | Percentage change of closing price over the last 1–5 candles |
| Volume_t, Volume_(t-1), Volume_(t-2), Volume_(t-3), Volume_(t-4) | Percentage change of volume over the last 1–5 candles |
| Body_size | Candle body size (Close − Open) |
| Range | Candle range (High − Low) |
| Body/Range_ratio | Ratio of candle body to its range |
| Upper_shadow | Upper shadow of the candle (High − max(Open, Close)) |
| Lower_shadow | Lower shadow of the candle (min(Open, Close) − Low) |
| RSI_14 | Relative Strength Index (14 periods) |
| MACD, MACD_signal, MACD_diff | MACD moving averages (12,26,9) and difference with signal line |
| ATR_14 | Average True Range over 14 candles |
| BB_bbm, BB_bbh, BB_bbl, BB_bbw, BB_percent | Bollinger Bands (middle, upper, lower, width, and percent position) |
| EMA_10, EMA_50, EMA_diff, EMA_200 | Exponential Moving Averages (10, 50, 200) and their difference |
| Distance_to_High, Distance_to_Low | Distance to local highs and lows over the last 48 candles (~1 day) |
| ATR_ratio | Ratio of current ATR to median ATR over 30 candles |
| OBV, OBV_slope | On-Balance Volume (OBV) and its slope (change) |
-
Fetch raw OHLCV data from Bybit API using:
data\raw_data\get_raw_candle.py -
Calculate technical indicators and prepare dataset with:
data\ready_data\processing_raw_candle.py
Output dataset is saved to:
data/ready_data/samples/dataset.csv
Normalized data is fed to the input of the neural network.
The models perform the same functionality, have similar performance parameters, but have different structures.
Model A: models\MLP\best_model_1
Model B: models\MLP\best_model_2
| Model A | Model B | |
|---|---|---|
| Number of layers | 4 | 3 |
| Number of input parameters | 34 | 34 |
| The number of neurons in the first hidden layer | 416 | 992 |
| The number of neurons in the second hidden layer | 16 | - |
| Hidden layer activation function | relu | relu |
| Number of neurons in the output layer | 3 | 3 |
| Output layer activation function | softmax | softmax |
| Optimizer | Adam | Adam |
| Loss function | Categorical Crossentropy | Categorical Crossentropy |
Model A: models\MLP\best_model_1
Model B: models\MLP\best_model_2
Model A:
Accuracy / Loss:
Confusion matrix:
Distribution:
Class distribution (y_test): [4708 4865 4988]
Prediction distribution: [4483 5133 4945]
Receiver Operating Characteristic:
Classification report:
Model B:
Accuracy / Loss:
Confusion matrix:
Distribution:
Class distribution (y_test): [4708 4865 4988]
Prediction distribution: [4400 5208 4953]
Receiver Operating Characteristic:
Classification report:
Options for interacting with the model:
1.Use console application:
git clone https://github.com/TalonProbeite/CryptoPriceML.git
cd CryptoPriceML
pip install -r requirements.txt
python src\usage\run.py