In this walk-through, we explain how Qraft was able to develop a market risk detection model, expressed through a cash allocation level, using artificial intelligence.

Key Points

Qraft has developed AI-powered market risk detection models designed to forecast weekly downside risk utilizing financial market data including momentum, volatility, correlation, and macroeconomic indicators
Qraft’s AI-powered market risk detection model has both backtested and live data that shows its success in avoiding severe drawdowns during major financial crises (GFC, US credit downgrade, 2020 COVID market crash) over the past 20+ years

Why Is This Important?

The rise in market volatility throughout the past few years has reinforced the importance of risk management. Not only has overall volatility increased over the past few years, but so has the magnitude of the volatility, as observed with tech stocks, with some falling more than 30% overnight.

This has been quantified with the changing trends in standard deviation over the past few years. According to Morningstar[1], volatility has risen sharply in 2022, with a standard deviation of 16.10% for the trailing 12-month period to April 30, 2022, contrasting with a standard deviation of 10.77% for the year 2021.

Avoiding dramatic drawdowns can have a significant impact on a portfolio, as drawdowns may take long to recover from, as the returns required to break-even increase exponentially with the magnitude of the drawdown (See Figure 1).

Figure 1. Table showing the returns needed to recover from drawdowns

Therefore, limiting drawdowns is at the forefront of the risk management philosophy, and a key research focus for Qraft. With the creation of a market risk detection model, we hope to help guide investors in assessing current market risk and helping them understand what that risk would mean to their individual portfolios. Additionally, we wanted to create a model that could be used directly by clients for portfolio risk management.

Qraft’s AI Regime Detection Model (monthly forecast model)

We hypothesized that a top-down macro approach would be needed, one that measured the volatility of the market environment at each cross-section of time. This would result in a model that could detect the current market regime, to provide a forecast of the immediate market environment.

This signal would then be converted to a corresponding cash allocation level, that would vary in response to the model’s assessment of the existing market risk.

The asset allocation of cash would therefore represent the level of volatility in the market, as predicted by the model. This meant that when cash allocation levels were high, predicted market volatility was high. A high asset allocation towards cash would then, in turn, completely or partially insulate an investor’s portfolio from possible drawdowns.

To maintain an active approach that could respond to market dynamism, it was decided a monthly update would be best, as most of the macro data used as inputs are released monthly. A monthly approach would work best to also preserve the integrity of the data, as a higher frequency approach would be difficult to implement using monthly released macro data inputs.

The resulting model, or the Qraft AI Regime Detection Model, wherein regime refers to the specific market regime that exists at a point in time. By using macroeconomic inputs to predict at the beginning of every month what the likelihood of the market regime (clusters of persistent market conditions) would be at a given month, the model would correspondingly allocate a portion of its assets in cash to avoid potential drawdowns.

The resulting portfolio was then compared to the general market, represented by the S&P 500. The light-gray bars in Figure 2 below represent the cash allocation for a given month on a scale from 0-100%. When overlayed onto a line graph showing the total return of the S&P 500 in the same period, we can visually see overlaps between the periods in which the AI Regime Detection Model has allocated a significant percentage of its assets to cash with periods of financial crisis or significant drawdowns.

Figure 2. Graph showing SPY performance overlayed with AI Regime Detection Model cash allocation levels

We then utilized the cash allocation signals from the model and then applied those allocation levels to a simulated fund based on the S&P 500, represented by the SPY ETF. The experiment found that while returns on our volatility-adjusted portfolio were behind that of SPY, because of the addition of dynamic cash allocation, our adjusted model performed much better on a risk-adjusted return basis, with a Sharpe ratio that is about 0.20 higher than that of SPY’s.

This model is currently being utilized by Hana Life Insurance, notably as the base of some of their variable insurance funds. Since its inception in 2019, and as of June 2022, the fund stood at approximately $200 million in AUM, delivering successful risk management strategies to its investors.

Refinement

While the monthly model has been successful, we have determined that the current market environment has been susceptible to increased volatility, with changes in market regime happening on a shorter frequency basis than monthly. Therefore, we set out to create a more that can dynamically respond to changing market conditions on a more frequent basis.

To refine this model, we decided to develop a new model with dynamic cash allocation available in various periods, including bi-weekly, weekly, and daily. This led to the creation of the AI MVC Model, which is a model that is currently being provided weekly.

Weekly AI Regime Detection Model (weekly forecast model)

To create a model for which dynamic cash allocation positions were available in periods other than the original monthly frequency, it was necessary to research to find new sets of input data that could be found in periods other than monthly and would yield similar results.

As alluded to earlier, options such as transposing existing macroeconomic data to fit into a weekly format with methods such as averaging, etc., would most probably not yield the best results as the data would not be completely accurate.

To get around this roadblock, an entirely different model was created, one we named the AI MVC (momentum, volatility, correlation) Model. We hypothesized that market data in the form of indices or ETFs that express market momentum and volatility, and the correlation between the two, would be able to be used as a substitute for macroeconomic data. This is because we believed the results of such macroeconomic data would have already been reflected within the movements of our selected indices and ETFs.

Momentum is used as a proxy for market direction, volatility is used to measure the scale of movement, and correlation is a proxy for the coupling effect. Using these factors, a three-dimensional prediction of the general market environment would be made.

Figure 3. Graph showing SPY performance overlayed with AI MVC Model cash allocation levels

The resulting MVC model has shown its ability to detect some of the major financial crises of the past decade and a half, as shown in Figure 3. above, where the light blue line represents the cash allocation levels set by the MVC model. We further analyzed the various financial crises that had occurred and the model actions during those periods and the results are shown in Figure 4.

Figure 4. Table showing AI MVC Model metrics during major financial crises

The AI MVC model was able to prevent drawdowns during major financial crises of the past 22 or so years, starting with the dot-com bubble crash. Through the table one can also observe when the model was able to first output a meaningful cash allocation level (allocation levels above 50), as well as the average cash allocation level of the model, during the crisis periods.

Figure 5. Table showing AI MVC Model performance metrics compared to that of SPY

This is further supported by the metrics shown in Figure 5, which makes it clear that Qraft AI MVC can deliver superior risk-adjusted returns as shown through the Sharpe ratio than that of the general market, represented by SPY.

Conclusion

We have demonstrated with our research how dynamic cash allocation levels can be utilized for risk management purposes. The Hana Life example mentioned previously, serves as a case study into the possibilities that exist with such a model and the demand for this type of risk management strategy. This is especially true given the rapid growth in the AUM of the fund, which now stands at around $200 million.

Another unique way this model is now being applied is through a partnership with Korea’s Maeil Business Newspaper. With this partnership, Qraft provides a variation of the AI MVC for Maeil’s AI Boom & Shock Index, which draws inspiration from CNN’s Fear and Greed Index. The AI Boom & Shock Index seeks to predict the market environment a week in advance. Its launch in May was met by large interest within the Korean finance industry, with verbal support from executives at major financial institutions within the country. These institutions have expressed interest in using this indicator/index as a solution or product for their clients. We have also recently launched the Qraft AI Risk On/Off Index, which while like the AI Boom & Shock Index, is a further refinement that is currently featured on Benzinga.

However, even with all these live-use cases, we believe this technology still holds much future promise, as we have already identified a few areas for potential improvement and growth. Variations of the model are also being explored with the potential considerations including replacing cash as an asset class with short-term bonds, reflecting the current inflationary environment.

[1] Arnott, Amy. “Why Market Volatility Is Nothing New.” Morningstar UK, Morningstar, Inc., https://www.morningstar.co.uk/uk/news/218728/why-market-volatility-is-nothing-new.aspx.

An AI Approach Towards Creating a Market Risk Detection Model