.GAIA:.The.Global.AI.Allocator - Systematic Investment Products and Decision Systems

GAIA: The Global AI Allocator WilmotML Dr. Aric Whitewood, Founding Partner

Opportunity and Approach “The future belongs to those who understand at a very deep level how to combine their unique expertise with what algorithms do best.” Pedro Domingos, The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World Systematic Investment Products Decision Systems AI Macro We implement systematic products using liquid, low-cost instruments such as futures and ETFs. Products are currently based on dynamic asset allocation rather than stock selection. Our approach to AI is that of human and machine intelligence combined. We design investment decision systems that are explainable.

Aspects of our Approach Established Macro Framework • We make use of an established macro framework, developed over the last 30 years, to guide aspects such as variable selection for the models. It also provides custom indicators such as Risk Appetite as inputs. AI and Deep Domain Knowledge The team has experience of applying data science and AI to financial market problems within investment banking and asset management. This includes dealing with: noisy data, changing regimes, and market behaviour. Flexible AI Platform • We have developed a flexible platform which makes evidence based predictions across markets worldwide. This can be used to power a variety of investment products, covering both asset (our initial focus) and security selection. Fundamental and Systematic • We blend fundamental insight (human intelligence) with very advanced data processing and systematic execution (machine intelligence). High Performance Computing • We use high performance computing techniques, which includes both GPUs and FPGAs. Explainable AI • We specialize in Explainable AI, which allows us to understand the prediction system, and also to explain its investment decisions to clients.

Machine Learning Approach ML Definition ML Advantages Efficient algorithms for searching different model specifications Diverse collection of high-dimensional models Regularisation methods for model selection and reducing overfitting High dimensional representations Produce complex, non-linear associations Enhanced flexibility Improved prediction quality ML Approach Staged design with transparency in mind Selection of techniques Curation of data

GAIA: The Global AI Allocator Outputs can be either discretionary or systematic Human PM/Trader Machine Driven Strategies Predictions for asset classes, sectors, factors: Decide how to adaptively allocate risk based on new information Diversity across multiple dimensions: representations, models, time Fundamental knowledge and macro framework guide choice of relevant input signals

Performance Best Practices: Our Approach Our research process adheres to certain best practices in terms of data and methodology, some aspects of which are shown below:

Glossary: Strategy Types and Performance Long-only / long-short • The first type of strategy, long-only, buys assets when it expects the price to rise, and sells when it expects the price to fall. It profits from upward movements in prices. • The second, long-short, buys in the same way, but goes short when it expects price to fall – we borrow the asset shares, then sell them, under expectation we can buy them again later when price is low. It so profits from both upwards and downwards movements in prices. Sharpe Ratio • This is a risk adjusted return, a performance measure for a strategy. Typically a Sharpe Ratio of 1 is good, 2 is excellent, and so on as it increases.

Diverse Representations As part of the Preprocessing stage, different representations of the input data are formed, based on transformations Thesetransformations may be simple or complex, and in some cases incorporate machine learning themselves A more diverse set of models is formed from this set of transformed inputs, which can be averaged, leading to improved performance. The example below shows the effect on Sharpe Ratio of averaging three different representations (producing the red line). Preprocessing Three Year Rolling Sharpe Ratio Distribution of Three Year Rolling Sharpe Ratio The effect of averaging is to produce Sharpe Ratios which have a smaller standard deviation – they are flatter over the entire period, whilst still maintaining a mean value of 2.4

Explainability Why? Our Approach Results Explanations for investors Regime Centric Key Drivers of Predictions Regime based models, but quite different from the existing literature. Also based on human function learning theory. Why did the machine make a particular decision? What decision is it likely to make next and why? Large number of models run for diversification. Reduced to key driving regimes and therefore variables for a given prediction. Interrogations for team Staged Design Relationships learnt by System Team has ability to see more detailed parameter choices and system state. Transparency and explainability designed into system from the outset. Different stages deal with different aspects of the processing, each can be interrogated separately. Discretionary Investing

Predictions: As Human and Machine Inputs Predictions take the form of return distributions for the next period: typically 1-4 weeks, but can be extended to around 3 months. Example distribution 1: positive mean Systematic Investment Decision 1 Discretionary Investment Decision 1 Probabilistic Predictions Discretionary Investing Systematic Investing Systematic Investment Decision 2 Discretionary Investment Decision 2 Example distribution 2: negative mean, positive skew

Predictions: Assets We predict a number of different assets, sectors, factors, and country equity markets globally. Below gives an indicative list, with additional assets being continually added. Assets Sectors Factors US Large Cap Equities US Technology US Momentum Probabilistic Predictions US Small Cap Equities US Healthcare US Quality US High Yield Bonds US Consumer Staples US Value US Treasury Bonds US Consumer Discretionary World Momentum UK Equities … World Quality China Equities China Industrials World Value Japan Equities China Technology … … …

Sources of Alpha “Alpha is the ability to form expectations which are better than the market” R.J.Fuller There are three main sources of alpha, described below: Information • We combine fundamental, macro and investor sentiment data to create regimes, or information sets. These represent the context of the market. Model • Machine learning helps us deal with non-linear, continuously changing relationships much more effectively than traditional quantitative techniques. Behaviour • We use the Credit Suisse Risk Appetite Index to represent investor sentiment, and have created alternate versions based on natural language processing.

Tactical Alpha What is the source of tactical alpha? “Modern markets show considerable micro efficiency (for the reason that the minority who spot aberrations from micro efficiency can make money from those occurrences and, in doing so, they tend to wipe out any persistent inefficiencies). In no contradiction to the previous sentence, I had hypothesized considerable macro inefficiency, in the sense of long waves in the time series of aggregate indexes of security prices below and above various definitions of fundamental values.” Paul Samuelson in a private letter to Robert Shiller (2001) How do we identify and take advantage of tactical alpha? We define the concept of a market regime, which lasts for some period of months (long waves as mentioned above). Note that market inefficiencies are typically classified as: fundamental, technical, or calendar based. The macro inefficiencies we identify through the above approach result from non-linear covariations of macro, price and sentiment data, and so can be viewed as combinations of these three types (above).

Investment Strategies We are currently running three investment strategies on proprietary capital; two long-only strategies and one long-short. The charts on the subsequent slides give returns and three year rolling Sharpe Ratios. From January 2018 is live out of sample data for the long-only, trading commences in August 2018. From November 2018 is live out of sample data for the long-short, trading commences in January 2019.

Performance GAIA Long Only USA : 3Y Sharpe Ratio • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only Global : 3Y Sharpe Ratio • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the Global Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. Balanced Portfolio is a 50/50 Portfolio of MSCI World and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Short USA : 3Y Sharpe Ratio • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US long short Variant of GAIA, which consist of historical out-of-sample data from March 2008 to October 2018, live model data from November 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only USA: Backtest vs. Live Track The distribution of daily returns is compared for the backtest (1996-2018) and the live trading period (2018-present day), shown here with the distributions.

Performance GAIA Long Only USA: Random Portfolios The distribution of mean Sharpe Ratios (three year) for random portfolios is shown here. The actual strategy is located above the 99th percentile of this distribution.

Performance GAIA Long Only USA : MA Portfolio Weights Worries over US economic and profit growth, trade war, and hawkish comments from Fed, contributed to market downturn in October and December 2018 Portfolio allocation to risky assets reduced Portfolio allocation to risky assets reduced Key drivers: USA and European growth forecasts, commodities, and US Government Bond yields Key drivers: Asian growth forecasts, commodities, and US Government Bond yields • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only USA : Returns Reduction in portfolio allocation to risky assets at the two points shown give superior downside protection with GAIA strategy • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Example Positioning and Drivers (Mid-March) Position Regime / Conditions US Dollar down Long US Stock Market Energy Sector up Commodities up Medium US High Yield US Growth down Japan Growth down 0% Japan (USD) ISM New Orders down

FPGA Acceleration We use accelerated compute as part of the training process, either with GPUs or FPGAs. GPUs provide an order of magnitude speed up compared to a CPU, for the particular algorithms we focus on. An FPGA can potentially achieve another one or two orders of magnitude speed up on the GPU, for the specific algorithms used here. This has to be balanced against the increased complexity of development, difficulty in finding people with the right skills, and increased effort in maintaining the resultant code base. We run an FPGA implementation on Amazon Web Services (Xilinx), programmed with OpenCL and VHDL. Ensemble Learning

Reinforcement Learning Systematic Strategies We have been researching the application of Reinforcement Learning (RL) to the portfolio optimisation stage (translating expected return probability distributions to portfolio weights). By taking the outputs of the established machine learning based predictions, we are reducing the dimensionality of the data input to the RL algorithms. Of course, this introduces its own bias as well. Our current implementation is based on a variant of Rainbow, a state-of-the-art Deep Q-Network. We test algorithms in an in-house developed event-driven market environment, inspired by the OpenAI-gym protocol. Probabilistic Predictions

Summary of Fund Terms Contact Us: WilmotML (XAI Asset Management) Dr. Aric Whitewood Founding Partner and Head of AI / Machine Learning aric.whitewood@wilmotml.com +44 203 890 2700 Jonathan Wilmot Founding Partner and Head of Macro Research jonathan.wilmot@wilmotml.com +44 203 890 2506 Office Address WilmotML, Level39, One Canada Square, London E14 5AB

Appendix

Predictions: Probabilistic Predictions take the form of return distributions for the next period: typically 1-4 weeks, but can be extended to around 3 months. 8 Example in-sample relationship between predicted value and return distribution (subsequent two weeks), Russell 1000. 2004-2008. Positive returns Compounded return/% 0 Probabilistic Predictions Negative returns -8 Prediction 8 Example out-of-sample relationship between predicted value and return distribution (subsequent two weeks), Russell 1000. 2008-2018. Positive returns Compounded return/% 0 Negative returns -8 Prediction

Negative Skew Indicator

Investor Sentiment: Risk Appetite “Risk Aversion”, “Risk Appetite”, “Investor Confidence”, “Investor Sentiment” Most macroeconomic and asset-pricing models incorporate some assumptions or model of Risk Appetite. Low Risk Appetite higher cost of capital, potentially limits business investment High Risk Appetite can produce booms in asset prices, and lead to recession A Brief Survey of Risk Appetite Indexes, Illing and Aaron

Investor Sentiment: Risk Appetite

Investor Sentiment: NLP Natural Language Processing (NLP) is at the intersection of these three fields (right). It can be split into three main areas (below): Speech Recognition Natural Language Understanding Natural Language Generation Similarity matching Calculate overall sentiment Rule based sentiment analysis engine Domain relevant keywords

Performance GAIA Long Only USA: Returns • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only USA : Relative to Balanced Outperformance is characterised by: Larger jumps in periods of market stress Gradual outperformance in other periods 2 1 2 1 The machine learning strategies can be thought of as continuously re-optimised balanced portfolios between sets of risky and safe assets. This chart shows relative performance of the GAIA strategy compared to a 50:50 balanced portfolio. 1 • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only Global: Returns • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the Global Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. Balanced Portfolio is a 50/50 Portfolio of MSCI World and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Only Global : Relative to Balanced As before, this chart shows relative performance of the GAIA strategy compared to an equivalent balanced portfolio. • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the Global Variant of GAIA, which consist of historical out-of-sample data from March 2008 to December 2017, live model data from January 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. Balanced Portfolio is a Portfolio of 40% MSCI World, 5% Commodities, 5% HY Bonds, and 50% US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Performance GAIA Long Short USA: Returns • Source: WilmotML/XAI Asset Management 2019 • The performance returns provided above are for the US long short Variant of GAIA, which consist of historical out-of-sample data from March 2008 to October 2018, live model data from November 2018. Transaction costs of 1.0% per annum are assumed. The unleveraged version is shown here, leverage up to 2x may be applied. USA Balanced Portfolio is a 50/50 Portfolio of Russell 1000 and US 10 Year Treasuries. They are neither indicators nor guarantees of future returns. Please refer to the important information regarding hypothetical, back-tested, projected or simulated performance at the end of this document.

Disclaimer The information contained in this presentation (the “Information”) is provided by XAI Asset Management Limited (the “Company”) to you solely for your reference. The Information has not been independently verified and may not contain, and you may not rely on this presentation as providing, all material information concerning the condition (financial or other), earnings, business affairs, business prospects, properties or results of operations of the Company or its subsidiaries. None of the Company or any of their members, directors, officers, employees or affiliates nor any other person accepts any liability (in negligence, or otherwise) whatsoever for any loss howsoever arising (including, without limitation for any claim, proceedings, action, suits, losses, expenses, damages or costs) from any use of this presentation or its contents or otherwise arising in connection therewith. This presentation contains statements that constitute forward-looking statements which involve risks and uncertainties. These statements include descriptions regarding the intent, belief or current expectations of the Company with respect to the consolidated results of the macroeconomic conditions. These statements can be recognised by the use of words such as “believes”, “expects”, “anticipates”, “intends”, “plans”, “foresees”, “will”, “estimates”, “projects”, or words of similar meaning. Similarly, statements that describe the Company’s objectives, plans or goals also are forward-looking statements. All such forward-looking statements do not guarantee future performance and actual results may differ materially from those in the forward-looking statements as a result of various factors and assumptions. You are cautioned not to place undue reliance on these forward-looking statements, which are based on the current view of the management of the Company on future events. The Company does not undertake to revise forward-looking statements to reflect future events or circumstances. No assurance can be given that future events will occur, that projections will be achieved, or that the Company’s assumptions are correct. Some statements, pictures and analysis in this presentation are for demonstration and illustrative purposes only. Any hypothetical illustrations, forecasts and estimates contained in this presentation are forward-looking statements and are based on assumptions. Hypothetical illustrations are necessarily speculative in nature and it can be expected that some or all of the assumptions underlying the hypothetical illustrations will not materialise or will vary significantly from actual results. No representation is made that any returns indicated will be achieved. Accordingly, the hypothetical illustrations are only an estimate and the Company assumes no duty to revise any forward-looking statement. This presentation may also contain historical market data; however, historical market trends are not reliable indicators of future market behaviour. Some statements and analysis in this presentation and some examples provided are based upon or derived from the hypothetical performance of models developed by the Company. Models are inherently imperfect and there is no assurance that any returns or other figures indicated in this presentation and derived from such models will be achieved. The Company expressly disclaims any responsibility for (i) the accuracy of the models or estimates used in deriving the analyses, (ii) any errors or omissions in computing or disseminating the analyses or (iii) any uses to which the analyses are put. By accepting and/or viewing the Information, you agree to be bound by the foregoing limitations.

.GAIA:.The.Global.AI.Allocator - Systematic Investment Products and Decision Systems