Finance 3.0: Uncertainty Management

Talos the Automaton:

Artificial intelligence and its guiding principles of evolution can be traced all the way back to the antiquity trough the Greek myths of Hephaestus and Pygmalion. Back in ancient Greece, storytellers’ orated epics about the intelligent robot and bronze giant Talos. As an automaton, the purpose of Talos was to protect Europa, the mother of king Minos of Crete, from foreign invaders. An automaton is a self-operating machine designed to follow a predefined set of rules fitting hand in glove with the definition of a quantitative trading system. The obvious difference between Talos and a systematic trading system is the objective function as the goal of a trading system is to generate high risk adjusted returns instead of protecting Europa from being assassinated. However, for Talos and the trading system to achieve their respective objective both of them must have a built in risk management system. Talos must differentiate friend from foe to prevent the wrong person from getting killed and the trading system must constantly balance reward against risk to avoid blowing up. Given recent developments in finance 3.0, the most advanced systematic trading systems now not only follow a pre-defined set of rules, but these superior systems now also recursively update their forecasting parameters to make enhanced predictions in the market. These adaptive quant systems are generally more complex than their finance 2.0 counterparts and to earn a higher risk adjusted return risk management is therefore more important than ever in modern quantitative investing.

Talos protecting Europa!

Uncertainty Management in Finance 3.0

Despite that the foundational mathematical theory of AI, spearheaded by cognitive psychologist (computer scientist) Geoffrey Hinton among others, have been around for over 50 years, it is recent advances in computing power and data gathering that are the main reasons for why there is now successful implementation of Machine Learning (ML) to financial markets. Through careful application of ML algorithms, high quality researchers combine a computer’s capacity to analyze data in infinitum with the dynamic ML model’s super human creativity to make optimal trade decisions. The ML model’s creativity is derived from recursive and complex mathematics to identify none-linear patterns in the financial markets. Whenever there is none-linearity involved the relationships between variables tend to move past the imagination of the human mind and when conducting abstract analysis a robust risk management system is extremely important to prevent the model from going haywire. For example, Adaptive Alph assumes that most of you readers would not want to sit in an AI powered car without a 99.999% guarantee of not crashing. Well, the AI investment process involves managing money, such as a worker’s future pension, and financial models must therefore equal the prudence of self-driving cars.

Geoffrey Hinton is known for his work on Artificial Neural Network dividing time between University of Toronto and Google.

Risk vs Uncertainty

To create a prudent risk management system for a quantitative investment portfolio, we must first understand the meaning of risk management. Adaptive Alph’s definition of risk management is the practice of identifying potential risks in advance, analyzing them and most of all taking precautionary steps to curb the risks that are identified. There are basically two components to risk causing negative performance. The first component is the probability of a risky event occurring; for example, the probability of a recession. The second component is the performance impact of a risky event; for example, the depth of a drawdown if a recession occurs. Nassim Talib provides excellent illustrations of risk in his must read book, Black Swan, which Adaptive Alph highly recommends if you want to understand the complexities of risk. Nassim has many points in his amazing book, but maybe his most important point is the difference between risk and uncertainty. Risk is when the outcome probabilities are known in advance. For example, the chance of winning by betting on a black or a red number in roulette is around 49.8% as the casino added the green 0 to skew the odds to casino’s advantage. Playing roulette at the casino many times therefore guarantees a 100% probability of going bust. One conclusion from the casino example is that the probability of a black swan is 0% because risk is measurable. However, risk or rather uncertainty in the financial market is inherently different from risk at the casino as black swan events can strike whenever and the potential impact may be unimaginable.

Roulette lacks uncertainty as all probabilities are known

LTCM

An unimaginable black swan event took place in 1998 when a group of genius “egg heads” at LTCM built models to capture arbitrage in the fixed income market. At first LTCM was the hottest hedge fund in the market and everyone wanted to invest as LTCM’s models kept racking up profits. To capitalize on arbitrage in the bond market, the LTCM models relied on capturing tiny anomalies in prices and to make money LTCM therefore needed to apply leverage on their positions. By applying leverage, LTCM essentially shorted volatility hoping that bond volatility remained within a predictable range. Myron Scholes coined the term “picking up nickels” and investors kept investing as they loved both LTCM and Myron’s elegant nickel analogy. However, picking up these nickels by applying leverage ended up being like a 100m sprinter using steroids to win the Olympic gold and then getting caught. LTCM’s risk management system completely underestimated their complex product, which contained many hidden risks that were also magnified by leverage. As a result, LTCM lost 50% of its value in August 1998 almost causing a global financial crisis. Due to LTCM’s popularity among investors, the fund was “to big to fail” so the central bank bailed them out to prevent a crash. In the end, LTCM is a black swan per Nassim’s definition because, 1) a ruin event took place when Russia devalued their currency and, 2) the event carried extreme impact. Risk management is therefore key for quantitative funds to prevent future LTCMs from happening. Given the low volatility over the past 10 years, I have a feeling that we have some current “eggheads” in the making. So what is then the solution for a quantitative hedge fund to protect itself from black swan type of events?

This book is awesome! It talks about LTCM and how one quant firm almost took down an entire economy

Risk Management: Portfolio Level

A successful systematic investor must take risks to profit, but unwanted risk must at all costs be avoided. To protect against unexpected shocks, the portfolio should therefore integrate risk management in its investment process, especially when it comes to machine learning. Integrating risk management means diversifying the portfolio while incorporating various risk measures such as max drawdown, volatility, VaR, skew and kurtosis that both monitors the trading system and controls the portfolio’s risk exposure. An ML trading system in finance 3.0 differs slightly from a finance 2.0 trading system despite that both of them relies on similar alpha indicators such as technicals and fundamentals. The reason for the difference is that an ML system dynamically evolves as new information enters the models. That means that the coefficients applied to the alpha indicators updates to hopefully make more accurate forecasts. One technique to prevent the parameters of a dynamic model from over adjusting is regularization. Applying regularization to an ML model is a combination of art and science. Only expert researchers who fully understand the data set and the objective function are able to appropriately optimize the rate of change of the parameters in the model. It is important to note that most machine learning models will make decision based on complex patterns generated by the market. Sometimes it is impossible for our human brains to understand patterns that are none-linear so having a robust risk management system is more important than ever to protect against black swan events.

To avoid risk concentration in a portfolio it is important to:

· Diversify across markets and models

o When diversifying across markets and models one accounts for correlation. If the correlation of a market or model is less than 1 to the existing portfolio and the expected return is greater than 0 then adding these to the portfolio will increase the risk adjusted return.

· Use spread positions

o Utilizing spread positions lowers the chance of taking a directional bet like LTCM. They used leverage on directional bets without considering volatility. A volatility that spikes causes margin calls.

· Lower exposures to those markets where the models fail to see opportunities.

o Usually systematic funds set a target volatility to generate an optimal risk adjusted return stream and to prevent extreme negative events from impacting the portfolio. The systematic investment process should therefore automatically reduce exposure to markets lacking opportunity where volatility increases to achieve the desired target volatility of the whole program. There are some disagreements among experts here.

Following are statistical approaches that can be integrated in the risk management process:

· Volatility

o Volatility is the most common. The benefit is that volatility is easy to understand and can be calculated using simple statistical methods. The weakness is that volatility does not care about the direction of the risk. You need volatility to make money. Higher downside volatility indicates a lower Sharpe.

· Max drawdown

o A maximum drawdown (MDD) is the maximum observed loss from a peak to a trough of a portfolio, before a new peak is attained. Maximum drawdown is an indicator of downside risk over a specified time period. MDD measures the size of drawdown, but it does not account for the frequency of losses.

The above graph demonstrates that as there are more trials the underlying distribution is more likely to follow a normal distribution. Despite the above tendency, Markets are not normally distributed.

Risk Management: Model Level

Before launching a quantitative portfolio of models, one must also incorporate risk management at the model level. There are basically two approaches to manage risk, the top-down and bottom-up approach. Most funds use a combination of both top-down and bottom-up in their risk management process. In a bottom-up approach, each model in the portfolio acts independently from one another. If model A sees an opportunity in market X it will increase the exposure to market X without consideration of model B. To prevent the risk exposure of being too large in a single market or asset class in the bottom-up approach, one could then use a top-down limit framework by setting VaR limits. VaR calculates the maximum loss expected (or worst case scenario) on an investment, over a given time and given a specified degree of confidence. Unlike volatility, VaR measures the worst-case scenario so it is directional. The fund will likely have predetermined VaR-limits on each market and the portfolio as whole. The weakness of VaR-limits is that they do not measure the size of the breach. This is a big weakness as in finance the return distribution of any market tends to not be normally distributed. One must therefore incorporate other risk measures such as the possibility of margin calls, kurtosis and skew. In a top-down approach, however, each model in the portfolio considers the exposure of another model in the portfolio to prevent concentrated risk exposures. Usually this is done through a constrained mean variance optimization. In the end, both risk management approaches uses diversification and statistical methods to optimize risk to achieve the highest return possible.

Usher is correct!

Conclusion

From the story of Talos, the bronze giant, we learned that the idea of AI has been around for a long period of time and thanks to advances in computing power researchers can now successfully build dynamic models that adapt to the evolving economy. These models capture none-linear patterns in the market and are difficult for our human brain to comprehend. As a model increases in complexity risk management becomes more important and finance 3.0 models therefore require a more robust risk management system than the typical finance 2.0 model. A successful risk management system uses a combination of statistical measures such as volatility, max drawdown and VaR at both the portfolio and model level to protect the investment portfolio. The return distributions of markets are not normal so one must also account for skew and kurtosis. There are two risk management approaches, the top-down and bottom-up approach. In the bottom-up approach the models act independently and in the top-down approach a portfolio optimizer is used to account for correlation between the models. Finally, we learned that there is a difference between risk and uncertainty through Talib’s casino example.

//Stay Adaptive!!