Predict altcoin's max price reach time and value using neural network based on bitcoin exchange market data.

ubuntu 16.04 (I didn't test with other OS / Version.)

Python3.6

Mysql 5.7+

Python libs in requirements.txt

The main purpose is making neural network model to predict the max price reach time from altcoin exchange market.

Let's see one of the altcoin price time graph.

Based on 12 hours history data, the NN model will predict "max price reach time index" and "max price value" in future 6 hours.

If we can predict what time altcoin price can reach the max value during specific times and how much price will be then it means we can make a money based on this NN model.

First, clone this repository.

git clone https://github.com/SkyHenryk/altcoin_max_price_prediction.git

Second, write mysql credentials information in "mysql_credentials.yaml".

host:
port:
id:
pw:
schema: quant_db

Third, get training data during 24 hours(1 day). It will get Bittrex trade data every minute.

After 24 hours, it will prepare training data based on trade data.

cd altcoin_max_price_prediction
bash build/get_training_data.sh

Fourth, start to training the neural network model based on prepared training data.

bash build/train_model.sh

Done! You can find the model result in "model_info" table in "quant_db" schema in Mysql.

Or, You can check the graph using tensorboard.

ex)

tensorboard --logdir ./tools/result/TrainerA_1/TrainerA_1_0.02_2500_10_0.2_32_300_relu_sigmoid_360_a1_last_max_index/graph
tensorboard --logdir ./tools/result/TrainerA_1/TrainerA_1_0.02_250_10_0.2_32_300_relu_sigmoid_360_a1_last_max/graph

RNN can predict the sequential value "right after" input sequential value well. But in this case, it needs to predict around 360 steps (6 hours -> 360 mins) and calculate max price and max reach time index. So I would say it is hard to predict.

But if I use ANN(artificial neural network) then the model just needs to predict only two output - max time index and max price. So it means ANN model is easier to reach higher accuracy than RNN model in this case. That's why I used ANN.