Creating a Multi-Strategy Fund Using Big Data and Machine Learning

  • By Paul Wilcox
  • December, 23
Blog Creating a Multi-Strategy Fund Using Big Data and Machine Learning



Constructing a Diversified Portfolio from Big Data to Maximize Risk-Adjusted Returns


Over the past year, I have written extensively on the benefits of a multi-strategy investment approach. A diversified, long/short portfolio of uncorrelated strategies is designed to profit in both risk-on and risk-off regimes.

Our inspiration for a multi-strategy fund stemmed from observations of our Model Portfolios outperforming their benchmarks in different periods of time. In many instances, we’ve witnessed one model dominate in outperformance, only to see a different one assume the lead soon after.

What is a Model Portfolio? Lucena’s Model Portfolios are perpetually traded simulations of an algorithmic approach to investment, powered by big data and AI.

The concept is rather simple. Backtests alone are notoriously plagued with biases, making their results subject to skepticism. In turn, we have resorted to Model Portfolios in which we carry forward the very same execution rules of a backtest perpetually into the future. All trades are determined algorithmically (no human discretion) and are published well before the market opens. 

Our goal is to simulate how AI models derived from one or more datasets can be deployed in real life scenarios. We apply best practices to remove the potential of presenting our simulations in an unrealistically favorable light. All backtests are performed out of sample, and both our backtests and Model Portfolios adhere to the following best practices:

a. Consider slippage, and transaction costs

b. Apply short borrowing cost, when applicable

c. Account for dividends, splits, and reverse splits

d. Incorporate flexible exit criteria such as conditional holds, OCOs (order cancel orders), stop losses, and target gains

e. Reduce path dependency with advanced allocation guidelines such as:

          – Max allocation per day

– Min/Max allocation per position

– Max number of positions per day

Ultimately, the purpose of a Model Portfolio is to help validate new investment ideas with empirical and defensible evidence before risking capital. Similar to Lucena’s backtest engine, the Model Portfolio results are available via a comprehensive performance analysis report.


Creating a Multi-Strategy Fund: Fab-5

By combining five of our top performing strategies into a long/short fund, we’ve in essence created a diversified low vol and high Sharpe portfolio. We call it the Fabulous 5. 

The Fab-5 is a collection of Lucena’s best quantitative research, geared to leverage the best sources of predictive data with the latest innovations in machine learning technology.  The fund enables each strategy to operate independently, but allocates its available cash based on a systematic mean-variance optimization (MVO) technology geared to maximize Sharpe ratio or risk adjusted returns.

Being a long/short fund, it combines positions from the five strategies as follows: two long only, two short only and one conservative long only portfolio of high-dividend utility and energy stocks. The underlying strategies are further described below:

Analyst Consensus – a long only strategy that identifies constituents from the Russell 1000 with “strong buy” analyst consensus. The strategy times and sub-selects its positions based on Lucena’s ranked technical and fundamental factors.

Dynamic Short Only – a combination of seven bearish scans. All scans participate in an ensemble set to identify high-conviction short entries. 

Utilities Live – consists of up to eleven high dividend yield securities that together minimize tracking error against the XLU (Utilities SPDR ETF). The portfolio re-optimizes its allocations periodically in order to maximize its Sharpe Ratio projections based on recent performance history.

Pre-Earnings Long – predicated on Wall Street Horizon’s earnings date restatement event. Lucena further hones in on additional fundamental factors (such as free cash flow, asset turnover, etc.) set to identify companies with positive outlook into their earnings release dates.

Pre-Earnings Short – predicated on Wall Street Horizon’s earnings date postponement event. Lucena further hones in on additional technical and fundamental factors set to identify companies with bearish outlook into their earnings release dates.


Correlation matrix

Correlation matrix of the 5 Model Portfolios. Shades of colors from Red (inverse correlated) to Green (fully correlated).


Fab-5 Perpetual Performance 

How is cash allocated between strategies in real time?

  • We view each strategy as a synthetic symbol which tracks the strategy’s performance over time. All synthetic symbols participate in a five-constituent portfolio that is optimized for a certain risk profile using mean variance optimization (MVO) with the following additional execution guidelines:

a. Apply a min/max allocation guidance of 1% to 25% per each synthetic symbol (i.e. strategy), both long and short

b. Re-optimize every three weeks — the optimizer relies on a three month lookback in order to determine the covariance between the strategies

c. Optimize target risk goal to most conservative (with the expectation that the fund will accrue returns slowly over time)

d. Account for transaction cost & slippage at a rate of 15bps round trip (7.5bps each leg) — in other words, we penalize our entry and exit prices by adjusting the published price against us

e. Apply stop loss and target gain guidelines at the strategy level while also employing a 6% stop loss per day at the fund level as a last defense circuit breaker for a rogue constituent or a black-swan event

  • Below you can see a backtest of the fund from 12/31/12 to 10/11/19 

Multi-strategy fund backtest

Out of sample, backtest of the combined Fab-5 fund against SP 500 benchmark. Past performance is not indicative of future results.

As can be seen, the Fab-5 fund outperformed the benchmark ($SPX) during the analysis period (12/31/12 to 10/11/19) by 56.76% while maintaining a lower drawdown of -9.12% vs. -19.8% and more than double in Sharpe ratio of 2.10  vs. 0.90 of the S&P 500.

The Model Portfolio was created in October. Below is the performance through 1/23/2020 the live report can be found here

Live Portfolio Using Big Data and Machine Learning

Past performance is not indicative of future results.


Maximizing the Value of Big Data and AI for Investment

Identifying uncorrelated best of breed strategies and combining them into a multi-strategy fund is not a new concept, it has been deployed by sophisticated hedge funds successfully for years. A new generation of sophisticated investment professionals are determined to take advantage of the compelling value of an algorithmic and diversified, low vol, high Sharpe portfolio that can dynamically adjust to changing market conditions.

At Lucena, we have automated the process of building such a fund from the ground up — validating alternative data sources, creating Model Portfolios powered by AI and big data, and ultimately combining uncorrelated models into a sustainable algorithmically traded market neutral multi-strategy fund. 


Questions about how we use AI to validate big data? Ask us below or contact us


Have a media inquiry or a topic you’d like to contribute to our blog?