Application of selected methods of statistical analysis and machine learning. learning in predictions of EURUSD, DAX and Ether prices

Transcription

1 Application of selected methods of statistical analysis and machine learning in predictions of EURUSD, DAX and Ether prices Mateusz Faculty of Mathematics and Information Science Warsaw University of Technology

2 Table of Contest Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network 4 Moving average crossover Logistic regression 5

3

4 Prices modeling Price modeling and prediction Trading strategies Application of novel ML and AI models Including new data sources in the models

5 Ecient market hypothesis[1] Weak-form eciency Future prices cannot be predicted by analyzing past prices Semi-strong-form eciency Neither fundamental analysis nor technical analysis techniques cannot be used in prediction Strong-form eciency Price reects also private information

6

7 Main prediction techniques Based on historical prices AR, MA, ARIMA models Bayesian networks Machine Learning Regressions k-nn SVM Articial Intelligence Social media (e.g Twitter, boards) Sentiment analysis

8 Using articial neural network models in stock market index prediction Figure: Using articial neural network models in stock market index prediction [5]

9 Using articial neural network models in stock market index prediction Related researches (goals): Comparing methods Exchange prediction Forecasting improvement Comparing ARIMA and ANN Crisis and bankruptcy prediction Investigate eect of volume on prediction Integration of fundamental and technical analysis Applying hybrid models

10 Neural networks performance in exchange rate prediction Figure: Neural networks performance in exchange rate prediction [4]

11 Neural networks performance in exchange rate prediction Main assumptions: Three-layer perceptron (5-10-1) is used EUR/USD, GBP/USD, USD/JPY Three steps Daily (83 values) Monthly (60 values) Quarterly (59 values)

12 Neural networks performance in exchange rate prediction Figure: NN model for EURUSD 1-day ahead prediciton [4]

13 Evaluating machine learning classication for nancial trading: An empirical approach Figure: Evaluating machine learning classication for nancial trading: An empirical approach [6]

14 Evaluating machine learning classication for nancial trading: An empirical approach Low complexity ML models USDJPY, EURGPB and EURUSD Multiagent system trading (2 years) 6 hour time trame Binary classication (up/down) Test set ( ) Test set II ( )

15 Evaluating machine learning classication for nancial trading: An empirical approach Figure: Results without retraining [6]

16 Evaluating machine learning classication for nancial trading: An empirical approach Figure: Results with retraining [6]

17 Evaluating machine learning classication for nancial trading: An empirical approach Figure: Results with retraining for best set up [6]

18 Price prediction is permanently open problem New data intervals and sources can be applied New models (hybrid) can be applied Market is not constant Results can be statistically insignicant

19 Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network

20 Ethereum graph Blockchain parsing Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network 1 Full blockchain download (via Parity) 2 Blockchain parsing to transaction list (via Parity API) 3 Transactions grouped into one-day packages 4 Few per-day measures extracted From to

21 Ethereum graph Ether price (in USD) Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Ether Price

22 Ethereum graph Number of nodes Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Number of nodes

23 Ethereum graph Number of edges Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Number of edges

24 Ethereum graph Total ow in Wei Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Total ow in Wei

25 Figure: Total ow in USD Ethereum graph Total ow in USD Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network

26 Ether correlations Time series Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Correlation of raw time series

27 Ether correlations Time series of returns Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Correlation of time series of returns

28 Ether correlations Lagged time series of returns Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Figure: Correlation of lagged time series of returns

29 Ether prediction model - Logistic Regression Model description Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Logistic Regression Dierences as features Rolling training/test set Training - 333, 498, 663, Test Results - AUC

30 Ether prediction model - Neural Network Model description Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Neural Network Activation - ReLu + sigmoid (last layer) Optimizer - Adam Loss - binary cross entropy MinMax scaler - based on train set Dierences as features Rolling training/validation/test set Training - 390, 490, 590,.. Validation Test Epochs from 50 to 2000 (500 when 20 inputs)

31 Ether prediction model - Neural Network Model variants Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Table: Model variants Model ethusd lag btcusd lag graph lag Neurons in HL Scaler Model N Model N Model Y Model Y Model N Model N Model N

32 Ether prediction model - Neural Network Results Ethereum (blockchain) graph Correlations Prediction model - Logistic Regression Prediction model - Neural Network Table: Results Model AUC.1 AUC.2 AUC.3 AUC.4 AUC.5 AUC.6 AUC.7 Model Model Model Model Model Model Model

33 Moving average crossover Logistic regression

34 Moving average crossover Moving average crossover Logistic regression Strategy based on 2 moving average cross - fast and slow Figure: Moving average crossover [2]

35 Moving average crossover Moving average crossover Logistic regression 5min interval divided into half Two indexes EURUSD DAX Long MA (SMA) lengths: 50, 100, Short MA (EMA) lengths: 5, 10,... 40

36 Moving average crossover EURUSD results Moving average crossover Logistic regression Figure: Mean return for long positions

37 Moving average crossover EURUSD results Moving average crossover Logistic regression Figure: Mean return for short positions

38 Moving average crossover DAX results Moving average crossover Logistic regression Figure: Mean return for long positions

39 Moving average crossover DAX results Moving average crossover Logistic regression Figure: Mean return for short positions

40 Logistic regression Moving average crossover Logistic regression 5min interval Two indexes EURUSD DAX %/25% - train/test split Prediction direction (up/down) 30min ahead

41 Logistic regression Moving average crossover Logistic regression Features Returns - r t, r t 1,...r t 5 Volume Log(Volume) RSI (n=14) WPR (n=14) MFI (n=14) 3 SMA deviation

42 Logistic regression EURUSD results Moving average crossover Logistic regression Figure: ROC (AUC = 0.536) for EURUSD

43 Logistic regression EURUSD results Moving average crossover Logistic regression Figure: Mean return for EURUSD test set

44 Logistic regression DAX results Moving average crossover Logistic regression Figure: ROC (AUC = 0.512) for DAX

45 Logistic regression DAX results Moving average crossover Logistic regression Figure: Mean return for DAX test set

46

47 Reliable price prediction is challenging problem Ether price Relationships between price and blockchain structure exists Using them in price prediction is still open question

48 Bibliography [1] E. E. Peters Teoria Chaosu a rynku kapitaªowe, 2005 [2] [3] Thien Hai Nguyen et al. Sentiment analysis on social media for stock movement prediction, 2015 [4] S. Galeshchuk Neural networks performance in exchange rate prediction, Neurocomputing (2015) [5] E.Guresen et al. Using articial neural network models in stock market index prediction, 2011 [6] E. A. Gerlein et al. Evaluating machine learning classication for nancial trading: An empirical approach, 2016