deep learning for long-term value investing€¦ · company structure joint venture between ‣...
TRANSCRIPT
DEEP LEARNING FOR LONG-TERM VALUE INVESTING
Jonathan MasciCo-Founder of NNAISENSE
General Manager at Quantenstein
COMPANY STRUCTUREJoint Venture between
‣ Asset manager since 1994
‣ Value philosophy
‣ Funds outperform on the long run
‣ AuM 3.7bn EUR (Feb. 2017)
‣ Large-scale NN solutions for superhuman perception and motor control
‣ ultimate goal of marketing AGI ‣ leverages 25-year track record
of IDSIA, one of the leading research teams in AI: ‣ recipient of the NVIDIA AI
pioneers award
EVOLUTIONARY-RL DEMO
Learned behavior from driver perspectiveLearned parking behavior at NIPS conference
RL to the real world Without a teacher, no supervision
WHAT WE DO
‣ Fully automated portfolio manager
‣ Long-Term Vision
‣ Build custom portfolios directly from fundamental data
‣ No human in the loop:
‣ Deep Learning and Reinforcement Learning
‣ Less biased
MAJOR DIFFERENCES BETWEEN FINANCE AND OTHER DOMAINS
‣ Rules of the game change over time: how to avoid forgetting what worked and not mixing things up?
‣ Lot of “state aliasing”: similar market configurations lead to opposite developments, state is only partially observable
‣ Limited history, and only one history
‣ No clear single objective, not as simple as classifying cats and dogs
‣ Rules for neural network design don’t transfer to finance as straightforwardly as it may seem
KEEP A LONG-TERM
VIEW ON THINGS
STOCK PICKER MODELS
SINGLE INSTRUMENT
DATABASE OF FUNDAMENTAL DATA
AI ALPHA GENERATOR
SUPERVISED SIGNAL
PREPROCESSING
‣ LSTM, CNN, etc.
‣ What supervised signal to use, and how to optimize for it?
ARE WE REALLY IN THE BIG DATA REGIME?
▸ Data: 10K companies, 20 years, new signal every month
▸ 240 data points per company, 2.4M data points in total
▸ Using sequences reduces the number of samples, what’s the sweet spot?
▸ Only one history and the rules of the game change over time
▸ Data augmentation:
▸ If good prior, one can try to augment the training data
▸ In finance if you have a good prior you don’t need AI
WALK-FORWARD TESTING
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
TrainingTesting
BACKTESTINGExpected
Real
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
WFT Step 1
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
TRA
INTE
ST
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
WFT Step 2TR
AIN
TEST
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
100
150
200
250
300
1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 2016
WFT Step 5TR
AIN
TEST
WALK FORWARD TESTING
▸ Tries to minimize “double dipping” as much as possible
▸ Can involve training a very large number of models
▸ e.g. monthly retraining for 10 years produces 120 training stages
▸ Tradeoff between retraining periods and target horizon not easy to determine, many models will have to tick at different time-scales
PLENTY OF DATA WHEN GOING END-TO-END
▸ Given a set of companies and their corresponding series of fundamental data produce a set of portfolios, optimized over a given time horizon, that maximize criteria such as SharpeRatio and InformationRatio
▸ Select a random start date
▸ Select a sub-universe of K companies out of the N
▸ this gets us a choose(K, N)-fold increase in the amount of data
▸ Issue with current systems is that they try to get alpha from fundamental data, what we want is conditional alpha. No prior on what is a good signal to be extracted, the system implicitly learns features that work for portfolio construction. This is the foundation of Deep Learning
No supervision on what signal to extract
DATABASE OF FUNDAMENTAL DATA
AI ALPHA GENERATOR
FEATURES0
PREPROCESSING
AI ALPHA GENERATOR
FEATURESN
PREPROCESSING
AI PORTFOLIO BUILDER
FEATURESNFEATURES0
Universe of companies
RISK
CONSTRAINTS
LOSS
Optimized portfolio
Company 0 Company N
SYSTEM TRAINING
EXPERIMENT CONFIGURATION
MANAGER
EXPERIMENT INSTANCE
GPU#0
EXPERIMENT INSTANCE
GPU#1
EXPERIMENT INSTANCE
GPU#N
RESU
LTS D
ATAB
ASE
FRONTEND REPORTING AND
ANALYSIS
Each EXPERIMENT INSTANCE runs a full WFT training
Pool of experimentsscales linearly with numberof GPUs, but no speedupfor single experiment
EXPERIMENT
WFT STEP 0
GPU#0
WFT STEP 1
GPU#1
WFT STEP T
GPU#N
GATH
ER R
ESUL
TS AN
D PA
CK TH
EM IN
TO R
ESUL
T OBJ
ECT
FRONTEND REPORTING AND
ANALYSIS
Each WFT step runs on a separate GPU in a MAP-REDUCE fashion
Experiment execution scales linearly with numberof GPUs.
RESULTS ANALYSIS AND VISUALIZATION
Outperformance Heat Map
Cumulative Performance
Rolling Performance
Performance Heat Map
BAYERNINVEST ACATIS KI AKTIEN GLOBAL MSCI WORLD INDEX
#positions 50 1654Performance 251.4% 104.7%
Performance p.a. 12.0% 6.7%Volatility p.a. 13.9% 13.0%
Return/Volatility 0.9 0.5Outperformance p.a. 5.3% —
Information Ratio 1.0 —Maximum Drawdown -49.1% -48.5%
Dividend yield 12M 2.5% 2.4%Calmar Ratio L36M 1.92 1.22
Investment Company BayernInvest, MünchenCustodian BayernLB, München
Manager ACATIS Investment GmbH, FrankfurtAI Model Developer Quantenstein GmbH, Frankfurt
ISIN DE000A2AMP25 (Institutional class)Bloomberg Ticker BIAKIAK GR Equity
Minimum Investment 50,000 Euro (institutional class)Investment Focus Equity Global
Domicile GermanyCurrency EUR
Benchmark MSCI World NDR (EUR)Inception March 23rd, 2017
Fiscal Year-End Dec. 31stFront End Fee Max 5%
Ongoing Costs 1.03%
Performance FeeAt present, starting at 3% outperformance 25% of yield generated by the fund during the settlement period is above the reference value MSCI World NDR (EUR).
Permission for Public Distribution DDistribution Distributed
MASTER FACTS
DISCLAIMER
▸ This document is only intended for information purposes. It is solely directed at professional clients or suitable counterparties in terms of the Securities Trading Act, and is not intended for distribution to retail customers.
▸ Past performance does not guarantee future results. Quantenstein accepts no liability that the market forecasts will be achieved. The information is based on carefully selected sources which Quantenstein deems to be reliable, but Quantenstein makes no guarantee as to its correctness, completeness or accuracy. Holdings and allocations may change. The opinions promote understanding of the investment process and are not intended as a recommendation to invest.
▸ The investment opportunity discussed in this document may be unsuitable for certain investors depending on their specific investment objectives and depending on their financial situation. Furthermore, this document does not constitute an offer to persons to whom it may not be distributed under the respectively prevailing laws.
▸ The information does not represent an offer nor an invitation to subscription for shares and is intended solely for informational purposes. Private individuals and non-institutional investors should not buy the funds directly. Please contact your financial adviser for additional information. The information may not be reproduced or distributed to other persons.
THANK YOU