Project MAGI

6/4/2016

Mizuho Artificial Generalized Intelligence

Sales Trading Department

Masahiko Todoriki

Copyright (c) Mizuho Securities Co., Ltd. All Rights Reserved.

Performance Improvement of Algorithmic Trading Strategies Using Deep Learning

1

1. Trading Algorithms

What are Algorithmic Trading Strategies

2

Buy 1500@376

Whenever there is a change in the market, the algorithm checks if the current situation fits the requirements to trigger executions.

Buy 1000@379 Buy 1100@378 Buy 900@379

Buy 1200@369 Buy 1100@363 Buy 900@363

Buy 1200@339 Buy 1100@349

TRADED QUANTITY / ORDER QUANTITY

AVERAGE TRADED PRICE

0/10000

0

1100/10000

349.0

2300/10000

343.8

3200/10000

349.2

4300/10000

352.7

5500/10000

356.3

6400/10000

359.5

7500/10000

362.2

8500/10000

364.2

10000/10000

365.9

9:08 9:36 10:07 10:44 11:22 12:50 13:33 14:16 15:00

DONE!!

An algorithm creates a rough schedule of trades such as “when”, “how many shares” and “at what price” and follow the schedule until all of its order quantity are traded. to buy or sell,

Fig1. A typical case of algorithmic trading

AI and Deep Learning

Machine Learning

Deep Learning Perceptron

Expert System

If Condition “A” Then Do Action “α” If Condition “B” Then Do Action “β” If Condition “C” Then Do Action “γ”

Copies how human “experts” would behave depending on specific condition

Basic

Fig2. Deep Learning on Trading Algorithms

Deep Convolutional Neural Network (CNN)

Advanced

Support Vector Machine (SVM)

Auto Encoders (AE)

Recurrent Neural Network (RNN)

Hidden Markov Model (HMM)

Reinforced Learning (RL)

Deep Belief Network (DBN)

DNN-HMM

Deep RL

DNN-RNN Existing Trading Algorithms

3

4

2. Our study of stock price prediction

What we predict

5

Predict the case when price of stock will have a significant change

+0.5% and above

from -0.5% to +0.5%

less than -0.5%

Prediction Time Spread 1 hour

Current Time 2 pm

Prediction Time 3 pm

Up

Down

Flat ±0.5%

Fig 3. Three Classifications of Stock Price Range at a Future Time

Threshold 0.5%

Dataset

6

Input Data

500 3200

Marketdata of Topix Core 30 constituents

Marketdata of Nikkei 225 futures

Recent 20 OHLC** + Volume • Minutely time series OHLCV*** (5 values) • 5-Minutely OHLCV (5 values) • Hourly OHLCV (5 values) • Daily OHLCV (5 values) • Weekly OHLCV (5 values)

100 most-recent order book data • Price and quantity of ask1 to ask8 (2 x 8

values) • Price and quantity of bid1 to bid 8 (2 x 8

values)

3200

100 most-recent trade data • Exec price from base price in % • Exec quantity vs, previous day total

traded volume in %

500

200

200

Label （Answer）

1 7800 Total

Fig 4. Structure of Input Data Used for Our Prediction

Type of Deep Learning Algorithm We Used

7

Hidden Layer 1

Hidden Layer 2

Hidden Layer 5

Hidden Layer 6

Output Layer

（4000） （3500） （2000） （1500）

Input Layer

（7800）

Fig 5. Structure of Deep Belief Network

Up

Down

Flat

Node 1

Node 2 Node 3

Node 4

Node 3997

Node 3998

Node 3999

Node 4000

Node 1

Node 2 Node 3

Node 3498

Node 3499

Node 3500

Node 1 Node 2 Node 3

Node 1998

Node 1999

Node 2000

Node 1

Node 2

Node 1499

Node 1500

Parameter 1

Parameter 2 Parameter 3

Parameter 4

Parameter 7797

Parameter 7798

Parameter 7799

Parameter 7800

Our Application

8

Throwing away the idea of creating one omnipotent AI

Ex. DBN1 is specifically trained to answer at 9 am predicting 10 am whether it’s in range between ±0.3% from current price or higher than that or lower than that DBN1

DBN2

DBN3 ・・・・・・

Fig 6. Create and Train Different DBN at Different Condition

Ex. DBN2 is specifically trained to answer at 1 pm predicting 1:30 pm whether it’s in range between ±0.15% from current price or higher than that or lower than that

Create many different DBNs for each specific conditions. (Current Time, Threshold and Prediction Time Spread)

Result

9

50%

55%

60%

65%

70%

75%

80%

85%

90%

95%

100%

900

910

920

930

940

950

1000

1010

1020

1030

1040

1050

1100

1110

1120

1230

1240

1250

1300

1310

1320

1330

1340

1350

1400

1410

1420

1430

1440

1450

DBN FREQAccuracy of our AI based prediction

Accuracy of prediction based on historical probability

Fig 7. Prediction Accuracy of Our AI Approach Time of day

+2.48% with low σ

Expected improvement of algorithmic trading strategy performance is 1 bps

10

3. Our business application

11

MAGI Platform Overview

What’s in MAGI

Ever evolving R&D platform to generate the best deep learning model which is specifically designed for market prediction!

Heterogeneous data sources are ready for training such as Historical Data( Stock, FX, Commodities), Financial Statements, News, and more…

1. Choice of AI

Provides common deep learning models such as DBN, RNN(LSTM), RNN(RBM), DNN-HMM.

2. Heterogeneous Data Sources 3. Easy to Train

Data preprocessing tasks and training tasks are schedules and run on multiple servers and on GPUs without programming!

Production Hardware of MAGI

12

Fig 8. Servers and Network

Infiniband Switch

Task Scheduler

Calculation Servers

224TFlops（NVIDIA Tesla M40 x 32）

Spec

Distributed Computing

Parallel File System + Raid50 Direct Memory Access

Low latency Infiniband 56Gbs network

System flow of MAGI

13

Set up training data

Schedule Server

CPU CPU GPGPU

Set up training purpose

Training

Prediction Result Database Prediction

Validation

Trained Networks

Preprocessing Progress Report

Training Progress Report

Algo/Trader/ Analyst/Quants

Performance Report

GPGPU Servers

Fig 9. System flow of MAGI

User Interface (GUI)

DB/Storage Users

Preprocessing distributed over CPUs on both Schedule and GPGPU Servers

Distribute training jobs to GPGPU

servers

Dispatches CUDA program

Refer

Set up training logic

14

Thank you for Listening！