cube Systematic Alpha Fund A Machine-Learning Based Systematic Fund

Transcription

1 cube Systematic Alpha Fund A Machine-Learning Based Systematic Fund

2 About Rcube Asset Management Rcube stands for Research, Returns and Risk management Started in 2011 as an investment research provider Currently managing two funds totaling $55 m AUM Rcube Global Macro Fund UCITS (launched in February 2014) Rcube Systematic Alpha Fund (launched in May 2015) Managers of the Rcube Systematic Alpha Fund: Paul Buigues, CIO Systematic Strategies pb@rcube.com Remi Takase, Fund manager rt@rcube.com 2

3 R C U B E Investment Process The case for Machine Learning-based trading

4 What is Machine Learning-based trading? Machine Learning Giving computers the ability to learn + = Systematic Trading Giving computers the ability to trade Machine Learning-based Trading Giving computers the ability to learn how to trade (and to trade accordingly) 4

5 The discretionary / systematic trading continuum 0% systematic Pure discretionary trading 25% systematic 50% systematic 75% systematic 95% systematic Discretionary trading with quantitative inputs Systematic with discretionary overrides (or vice versa) Traditional systematic trading Machine Learning-based trading Machine Learning: the most extreme systematic trading approach. 5

6 Traditional systematic trading 75% systematic Trading system Validated Trading System designer Backtest Execution Not validated System abandoned due to weak live performances Trading systems are the assets of traditional systematic trading. Trading those assets can be as emotional as discretionary trading. 6

7 Machine learning-based trading 95% systematic ML algorithms Data universe Selected TS Coder / data scientist Automated TS discovery & selection TS execution New market observations=> learning When ML is used, trading systems are discovered and selected by algorithms. Humans only provide computer code and data sets. 7

8 Man vs. Machine: Trading System discovery speed Manual discovery Automated discovery TS Code Manual validation 1 trading system / hour trading systems / hour 8

9 Performance Performance Man vs. Machine: validation accuracy Manual trading system selection Automated trading system selection Selection point Selection points Time Due to selection bias, a group of manually selected trading systems will underperform when traded live vs. backtests. Validation is therefore impossible (or at least misleading) Time A continuous and automated selection of trading systems enables forward testing (no selection bias) Forward test results are close to live performances (WYSIWYG) 9

10 Machine Learning: two alternative approaches A small number of complex models - Each model has many inputs - Models are based on non-parametric techniques such as Neural Networks: A large number of simple models - Each model only has a few inputs - Models are based on simple and robust statistical techniques Input 1 Input 2 Input 3 Input 4 Input 5 Input Output This approach is valid in fields where rules do not change over time. It generally does not work in finance, which is a complex adaptive system. Simple models are structurally much more persistent than complex ones. => We opted for this approach 10

11 Recap: Successful human traders vs. Machine Learning Successful traders Machine Learning Quick-witted Disciplined Unemotional Patient Thinks probabilistically Able to adapt Have a critical mind Under continuous improvement 11

12 R C U B E Investment Process Rcube Systematic Alpha Fund Overview

13 Quick facts Investment Strategy Systematic trading based on Machine Learning Target return 10% per year over the risk-free rate Target volatility range 10-15% annualized Investment Universe Highly liquid instruments (in equities, rates, forex, commodities) Fees (A shares) Management fees: 1.5% / performance fees: 15% Fund type Cayman fund Fund inception May 1 st, 2015 Strategy track record 3 months since launch (plus 14 months as a sub-portfolio) 13

14 Trading Systems The fund is a portfolio of thousands of trading systems involving 9 assets. For a given asset, trading systems are based on: Economic data: 80% Other assets prices / returns: 20% Prices/returns of the asset itself: <0.1% By design, a trading system has a neutral long-term average exposure to the assets it trades (no structural beta => pure alpha) 14

15 Investment process Step 1 Trading system generation & selection Step 2 Trading system allocation Step 3 Execution For each of the assets we trade, thousands of trading systems are selected Allocation to trading systems, subject to the portfolio s aggregate volatility target Daily execution of the aggregate portfolio of trading systems 15

16 Step 1: Trading system generation & selection For each of the 9 assets we trade, machine learning algos generate and select thousands of trading systems. S&P 500 Estoxx 50 MSCI EM US 10Y DE 10Y Dollar Index Gold Copper Crude Oil Examples of trading systems generated by machine learning algorithms:... Asset: S&P 500 Input: SLOS Survey Asset: Eurostoxx 50 Input: German 10Y yields Asset: Gold Input: Consumer Sentiment LONG Exposure to the asset SHORT In-sample performance and Sharpe ratio Sharpe = 0.44 Sharpe = 0.67 Sharpe =

17 Step 2: Trading system allocation 9 assets S&P 500 Estoxx 50 MSCI EM US 10Y DE 10Y Dollar Index Gold Copper Crude Oil Thousands of trading systems per asset Trading system aggregation S Trading systems are aggregated according to: - their past risk-adjusted performance - their expected persistence One aggregate trading system for each asset (9). The allocation to aggregate trading systems is subject to the portfolio s volatility target. 17

18 Step 3: Execution Every night, trading systems are selected and aggregated. At 7am CET, the system generates new target positions for each asset. After checking for potential errors, we trade the delta between current fund positions and target positions (using execution algos). Typical daily trading activity (for $ 25m): Asset code Current # of contracts Target # of contracts Delta ES1 Index VG1 Index MES1 Index TY1 Comdty RX1 Comdty DX1 Curncy GC1 Comdty CL1 Comdty HG1 Comdty We trade on average around 10% of the portfolio s nominal size every morning. Note: Trading the aggregation of thousands of trading systems vs. trading individual systems dramatically lowers trading costs 18

19 Forward test procedure Step 1 Trading system generation & selection Step 2 Trading system allocation Step 3 Execution Rerun the process at times t, t+1, t+2, T Computing power needed to forward test 200 assets since 2000: around 300 Trillion arithmetic operations 19

20 Risk management Risk management is embedded in the investment process (maximum volatility target: 15% annualized) Risk is also managed independently by our risk manager (using HedgeGuard Financial Software) If the NAV experiences a 10% drawdown, we cut the portfolio for a minimum period of one week. Discretionary interventions: When we believe that an asset s former drivers are not going to be relevant for a certain time, we can discretionarily adjust (or eliminate) the strategy s target position for the asset. This can take place in the event of: Major geopolitical events Changes in the nature of an asset (e.g.: pegging) 20

21 The importance of trading system diversification Correlations between beta-neutral trading systems are considerably lower than correlations between assets. Running a diversified set of trading systems vastly increases the Sharpe ratio of the portfolio. r=0 r=0.01 r=0.05 r=0.20 r=

22 R C U B E Investment Process Track record & Conclusion

23 Live track record - Sharpe ratio: 1.13 (net of fees) Annualized rate of return: 11.64% Annualized volatility: 10.26% Performance before May 2015 corresponds to the Systematic Alpha Program levered by 1.5x (12% volatility target). Management fees (1.5%) and performance fees (15%) are taken into account. Before May 2015, the Systematic Alpha Program was run as a sub-portfolio of the Rcube Global Macro Fund. 23

24 Forward test backfill (01/ /2014) Forward test of 90,000 trading systems on 9 assets (no selection nor look-ahead bias) Target volatility:12% 24

25 Rcube Systematic Alpha Fund: Conclusion An innovative investment process Harnessing the collective intelligence of tens of thousands of virtual traders 17 month live track record 10% target excess return with 10-15% volatility target range 25