TRADING USING DEEP LEARNING MAN VS MACHINE Orders By Algorithms - PowerPoint PPT Presentation

TRADING USING DEEP LEARNING

MAN VS MACHINE Orders By Algorithms 84% Orders By Human 16% TRADING USING DEEP LEARNING

Artificial Neural Networks Neural networks are a family of models inspired by biological brain structure and are used to estimate or approximate functions that can depend on a large number of inputs and are generally unknown. Recent breakthroughs in artificial neural networks led to a modern renascence in AI. TRADING USING DEEP LEARNING

DEEP LEARNING SUPERIORITY Deep Learning 96.92% Human 94.9% ref: http://www.image-net.org/challenges/LSVRC/ DEEP META LEARNING

GRADIENT DESCENT 𝑥 𝑢 = 𝑥 𝑢−1 − 𝛿 𝜖𝐹 𝜖𝑥 𝐹 = Error of the network 𝑋 = Weight matrix representing the filters DEEP META LEARNING

GRADIENT BASED MODELS 1: Forward Propagation 2: Loss Calculation Legend 3: Optimization 𝑧 𝐹 = 𝑚 ො 𝑧, 𝑧 𝑧 – Ground Truth 𝜖𝐹 = 𝜖𝑚 ො 𝑧, 𝑧 𝑦 0 - Features Vector 𝑔 𝑜 𝑦 𝑜 , 𝑥 𝑜 = ො 𝑧 𝜖𝑦 𝑜 𝜖𝑦 𝑜 𝑦 𝑗 - Output of 𝑗 layer 𝑥 𝑜 𝐶𝑏𝑑𝑙 𝑄𝑠𝑝𝑞𝑏𝑕𝑏𝑢𝑗𝑝𝑜 𝜖𝐹 = 𝜖𝐹 𝜖𝑔 𝑜 𝑦 𝑜−1 , 𝑥 𝑜 𝑔 𝑜−1 (𝑦 𝑜−1 , 𝑥 𝑜−1 ) 𝑥 𝑗 - Weights of 𝑗 layer 𝐺𝑝𝑠𝑥𝑏𝑠𝑒 𝑄𝑠𝑝𝑞𝑏𝑕𝑏𝑢𝑗𝑝𝑜 𝑥 𝑜−1 𝜖𝑦 𝑜−1 𝜖𝑦 𝑜 𝑦 𝑜−1 𝑧 – Model Output ො 𝑔 𝑜−2 (𝑦 𝑜−2 , 𝑥 𝑜−2 ) 𝜖𝐹 = 𝜖𝐹 𝜖𝑔 𝑜 𝑦 𝑜−1 , 𝑥 𝑜 𝑚 ො 𝑧, 𝑧 𝜖𝑥 𝑜 𝜖𝑦 𝑜 𝜖𝑥 𝑜 - Loss Function 𝐹 – Loss Surface 𝜖𝐹 𝜖𝐹 𝜖𝑔 𝑜−1 𝑦 𝑜−2 , 𝑥 𝑜−1 = 𝑔 2 (𝑦 2 , 𝑥 2 ) 𝜖𝑦 𝑜−2 𝜖𝑦 𝑜−1 𝑦 𝑜−2 𝑔 – Activation Function 𝑥 1 𝜖𝐹 𝜖𝐹 𝜖𝑔 𝑜 𝑦 𝑜−2 , 𝑥 𝑜−1 𝑔 1 (𝑦 1 , 𝑥 1 ) = 𝜖𝑥 𝑜−1 𝜖𝑦 𝑜−1 𝜖𝑥 𝑜−1 𝑥 0 𝑔 0 (𝑦 0 , 𝑥 0 ) … 𝑦 0 … DEEP META LEARNING

Learning From Examples. Supervised Learning in a nutshell. CAT

FINNANCIAL PREDICTION PITFALLS Importance Overfitting Behavior Data Importance is Overfitted easily, Behavior of financial questionable and most models have markets change all the determination of poor predictive time and can be really meaningful data is capabilities unpredictable. hard. On financial data. Much Data No Theory Noisy Data Possible relevant Complex non-linear Noise In financial data data from many interactions in the data Is very common and markets is incredibly are not well specified sometimes large. by financial theory. distinguishing noise from behavior is hard. TRADING USING DEEP LEARNING

WHY DEEP LEARNING? Importance Overfitting Behavior Data Importance is Overfitted easily, Behavior of financial questionable and most models have markets change all the determination of poor predictive time and can be really meaningful data is capabilities unpredictable. hard. On financial data. Much Data No Theory Noisy Data Possible relevant Complex non-linear Noise In financial data data from many interactions in the data Is very common and markets is incredibly are not well specified sometimes large. by financial theory. distinguishing noise from behavior is hard. TRADING USING DEEP LEARNING

.5 .5 .5 .5 1 2 3 4 35:09.9 37:08.7 38:59.2 40:38.7 160.5 161.5 162.5 163.5 42:17.6 161 162 163 164 43:49.0 45:35.9 35:09.9 47:55.2 36:57.2 49:32.2 38:04.2 51:02.2 40:03.3 52:25.6 41:06.0 54:24.2 42:41.7 56:31.8 44:12.7 58:27.3 45:35.9 59:58.8 47:39.1 01:43.3 49:05.9 03:22.2 50:17.6 08:03.3 51:31.1 10:11.6 53:04.0 14:04.0 BACK TO FINANCE 54:36.3 17:37.5 56:31.8 21:19.2 58:07.0 26:03.8 59:29.3 30:46.0 00:57.2 33:07.0 02:31.0 36:54.2 04:05.5 TRADING USING DEEP LEARNING 40:00.4 08:16.0 43:27.5 10:11.6 48:22.7 12:09.7 52:26.2 17:17.8 54:59.0 20:37.3 57:33.4 22:09.9 01:00.6 28:36.1 04:41.4 30:52.1 08:23.5 33:07.0 11:20.4 36:46.1 ? 15:44.1 39:29.7 19:15.9 42:05.1 21:24.3 46:26.1 25:00.7 49:55.8 30:25.5 53:08.2 37:26.4 54:59.0 43:34.5 57:12.2 47:56.6 00:35.6 53:40.0 04:10.1 01:46.4 06:56.1 21:12.6 09:30.2 31:31.7 12:35.7 44:02.0 15:44.1 59:33.3 19:03.8 12:50.0 21:05.0 23:13.6 28:12.2 32:33.9 38:22.9 43:34.5 46:47.4 52:54.0 00:33.2 09:54.6 22:43.6 33:06.1 44:02.0 59:27.2 03:38.7 14:35.6

Strategy Universe Strategy Configurations Configuration Trading Decision Utility 1 - buy 0 - hold P&L / Drawdown -1 - Sell DEEP REINFORCEMENT LEARNING

Technical Analysis Technical analysis Might of might not work, One thing for sure: Very hard to generalize. TA-Lib : Technical Analysis Library talib.SMA( … talib.MOM( … DEEP LEARNING IN FINANCE

Successful Technical Trading Agents Using Farnsworth , Genetic Programming 2004 Surprisingly, Genetic programing can be very successful when It comes to financial strategies gp = SymbolicRegressor(... gp.fit(X_train, y_train) DEEP LEARNING IN FINANCE

DEEP LEARNING IN FINANCE

DEEP APPLYING DEEP LEARNING TO ENHANCE LEARNING MOMENTUM TRADING STRATEGIES IN STOCKS IN FINANCE

APPLYING DEEP LEARNING TO ENHANCE MOMENTUM TRADING STRATEGIES IN L Takeuchi, STOCKS 2013 FEATURE ENGINEERI 12 Monthly Returns NG For every month: Daily cumulative returns Z-score Against other cumulative returns of other stocks Flag if January 𝐺𝑓𝑏𝑢𝑣𝑠𝑓 𝑊𝑓𝑑𝑢𝑝𝑠 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝑢+𝑗 𝑐𝑗𝑒 𝑢 − 𝑏𝑡𝑙 𝑢−12 𝑐𝑗𝑒 𝑢+1 − 𝑏𝑡𝑙 𝑢−12 𝑐𝑗𝑒𝑢−𝑏𝑡𝑙𝑢−12 𝑐𝑗𝑒𝑢−𝑏𝑡𝑙𝑢−12 −𝜈 −𝜈 𝑌: > ෍ 𝑌 𝑗 : σ 𝑢−12 | 𝑗 ∈ (1,11) , ∪ 1 𝑗𝑔 𝑢 𝑗𝑜 𝑘𝑏𝑜𝑣𝑏𝑠𝑧 𝑓𝑚𝑡𝑓 0 𝑌 𝑗 : σ 𝑢−12 | 𝑗 ∈ (1,11) , ∪ 1 𝑗𝑔 𝑢 𝑗𝑜 𝑘𝑏𝑜𝑣𝑏𝑠𝑧 𝑓𝑚𝑡𝑓 0 𝑢+𝑗 𝑏𝑡𝑙𝑢−12 𝑢+𝑗 𝑏𝑡𝑙𝑢−12 𝑏𝑡𝑙 𝑢−12 𝑏𝑡𝑙 𝑢−12 𝜏 𝜏 𝑢−12 MODEL K - Fold Stacked Not Written 33 40 4 50 33 2 40 RBMs Layers size found by grid search Hyper Parameters Structure Layers RESULT S Confusion Matrix Accuracy: 53.061% Predicted Predicted Precision: 61.224% True False True Recall: 53.659% 22.38% 27.45% False 19.19 % 30.97% DEEP LEARNING IN FINANCE

DEEP DEEP MODELING COMPLEX COUPLINGS WITHIN LEARNING FINANCIAL MARKETS IN FINANCE

DEEP MODELING COMPLEX COUPLINGS WITHIN FINANCIAL MARKETS FEATURE ENGINEERI Used Deep Belief Network to find hidden couplings between markets NG Used Past Prices of stocks and forex as features Unsupervised Learning Model 𝐺𝑓𝑏𝑢𝑣𝑠𝑓 𝑊𝑓𝑑𝑢𝑝𝑠 𝑌 𝑗 : 𝑇 𝑗 ∪ 𝐺 𝑗 S R MODEL T Loss: Negative Log B O C R R M Likelihood DBN of Stacked K B B Optimizer: SGD RBMs Note: No Cross Validation in F R M M O B R paper E M Hyper Parameters Structure X Layers RESULT S DEEP LEARNING IN FINANCE

DEEP DEEP LEARNING FOR MULTIVARIATE FINANCIAL LEARNING TIME SERIES IN FINANCE

DEEP LEARNING FOR MULTIVARIATE FINANCIAL TIME SERIES FEATURE ENGINEERI Matrix of log returns over all the stocks NG Z-score Against other stocks log returns Flag if January 𝐺𝑓𝑏𝑢𝑣𝑠𝑓 𝑊𝑓𝑑𝑢𝑝𝑠 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝑐𝑗𝑒𝑢 log −𝜈 | 𝑗 ∈ (−33, −2) , ∪ 1 𝑗𝑔 𝑢 𝑗𝑜 𝑘𝑏𝑜𝑣𝑏𝑠𝑧 𝑓𝑚𝑡𝑓 0 𝑍: 1 𝑗𝑔 𝑄𝑠𝑗𝑑𝑓 𝑗𝑡 𝑝𝑤𝑓𝑠 𝑛𝑓𝑒𝑗𝑏𝑜 𝑏𝑢 𝑢 + 1 𝑏𝑡𝑙𝑢−1 𝑌 𝑗 : 𝜏 “ Deep ” Belief Net Fully Connected MODEL Loss: Negative Log Likelihood DBN Connected to D R R R Regularization: 𝑀 1 E MLP B CrosVal: Tarining: 70%, Valid: 15%, Test: B B N Activation: 𝑢𝑏𝑜ℎ 1 S M M M Optimizer: ADAGrad 15% e Hyper Parameters Structure Layers RESULT S DEEP LEARNING IN FINANCE

DEEP IMPLEMENTING DEEP NEURAL NETWORKS FOR LEARNING FINANCIAL MARKET PREDICTION IN FINANCE

IMPLEMENTING DEEP NEURAL NETWORKS FOR FINANCIAL MARKET PREDICTION FEATURE ENGINEERI All moving averages from 5 to 100 NG List of 100 lagged prices Pearson correlation between the returns (all 100) of the stock and all the other stocks(45) 𝐺𝑓𝑏𝑢𝑣𝑠𝑓 𝑊𝑓𝑑𝑢𝑝𝑠 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝐻𝑠𝑝𝑣𝑜𝑒 𝑈𝑠𝑣𝑢ℎ 𝑗−100 𝑞 𝑢 − 𝑞 𝑢−1 100 𝑏 𝑌: ራ ∪ ራ 𝑁𝐵( 𝑌, 𝑔) ∪ ራ 𝜍(𝑌, 𝑌 𝑘 ) 𝑍 ∈ 1,0, −1 𝑔𝑝𝑠 𝑐𝑣𝑧, 𝑡𝑓𝑚𝑚, ℎ𝑝𝑚𝑒 𝑞 𝑢−1 𝑗 𝑔=5 𝑘=1 Fully Connected MODEL Loss: Categorical Cross Simple Fully Entropy CrosVal: Tarining: 80%, Test: Connected Activation: 100 135 1 9895 1000 ReLU, 𝑡𝑝𝑔𝑢𝑛𝑏𝑦 20% Training Algorithm: Walk Optimizer: SGD forward Hyper Parameters Structure Layers RESULT S Accuracy: 73% F1 Score: 0.4 DEEP LEARNING IN FINANCE

DEEP DEEP LEARNING IN FINANCE LEARNING IN FINANCE