Difficulty level:

How do you come up with new ideas for algos?
What do you look for when creating new algorithms?
What data do you use?
What does the process look like?
How long does it take to create a new algorithm?
Do you continually adjust the algorithms?

At Unhedged, our Founders have been writing algorithms for many years. The current Unhedged Algorithms have resulted from this process.

Algorithms are generally based on academic research. We constantly review market factors such as size, earnings, quality and momentum of stocks and bonds. This helps us to determine whether our existing algorithms are sufficiently uncorrelated or whether we can include new alternative data or triggers for specific market conditions. In this way, we decide if we need to update an existing algorithm or if we need to create an entirely new one.

Enter the Factor factory
The base of an algorithm is usually a factor. Factors can give an indication of market movements. Some common factors are:
Momentum (buy stuff that goes up),
Size (buy big stuff),
Mean Reversion (buy things that went up or down too fast),
Earnings improvement (buy stocks of which earnings are improving)
Gender equality (buy stocks where the board/mgt is balanced in gender),
R&D spend (buy companies that invest a lot in innovation: it will pay off)
Brand value (buy shiny things that have built valuable brands)
But there are many more!

Correlation is core to the performance of an existing algorithm and needs to be considered when building a new algorithm. Correlation indicates how similar the movement of returns are between algorithms. A combination of algorithms with low correlation will typically generate less volatile returns than one algorithm on its own.

When creating a new algorithm, we look for one that is either uncorrelated or only weakly correlated to the existing live algorithms. This means that we should not look into a factor that is already covered in an existing algorithm. This will increase correlation. A combination of algorithms that consider different market factors will typically be less volatile than one algorithm that focuses on one market factor.

At Unhedged, we develop algorithms that are not highly correlated (less than a 0.4 ρ correlation coefficient, often lower than 0.1), which, on average, creates lower maximum losses and delivers a less volatile return profile.

Alternative data
Data is super important! There are many types of data that can signal market movements. Our algorithms use historical pricing but we also look for alternative data to feed our algorithms so that we include information that the market may have missed.

For example, you can count cars in a parking lot of a retailer to predict their revenue. If you were able to work out the breakdown of vehicle makes and their value you could potentially go further and predict profits. In the same way you can also analyse company supply chains to determine whether there is a suggestion of a new product being developed upstream, or a significant issue that will affect downstream producers.

An Iterative Process

1. Get and clean the data
We collect our data from multiple sources and a key element for us is sourcing new and alternative data that may have been overlooked by the market. We ensure the data is clean for example, by ensuring no forward-looking biases exist and that it is accurate at the relevant "point-in-time". We then split the data into Developing, Testing and Forward Testing datasets because we want to make sure we have a different set to train and test our algorithms on.

2. Seek alpha, bear beta
Next, we try to isolate a clean "Alpha" factor on the development data set. (Alpha is a return that is independent of the market's returns, market returns are known as beta returns). When the consistency and statistical significance of our Alpha becomes sufficient, we stress test the Alpha in other asset classes or random datasets to see if we can get the same results.

Many High-Frequency Hedge funds do the opposite of what we do. They try to isolate the beta (the market returns) and aim to ‘cancel out’ market movements. We believe that we should use the market instead. It is a fundamentally different approach to algorithmic trading. Un-Hedged… Get it?

3. Come alpha, taste some actual data
The next step is to put the Alpha factor into a standard Algorithm without any guardrails e.g. no stop losses, then we run the algorithms on testing datasets and monitor its behaviour. Then we add some stop-loss and other filters and run the algorithm on a dataset that it's never seen before.

The next step is to trade on a paper account. If this is successful, that is backtesting and trading are close to identical, we expose it to real money and real-live data.... exciting and scary! When we say real money, we put the algo to work with minimum possible funds from the company (never customer’s funds). We then monitor it for a few weeks, months, even years until we are satisfied that the performance and statistical properties on real money are similar to the previous phases.

4. Ensure compliance
We then present the Algo and all the data to the Investment Committee and the Custodian for approval. When everyone's happy, we run the Regulatory compliance process adding the new Algo to our Important Documents (e.g. the Product Disclosure Statement) and once we have final approvals we can offer it to customers.

5. Monitor & improve
After launch, things are not 'done-and-dusted'. We continuously monitor all algorithms for statistical anomalies and feed the information back to the beginning of the process. Our quants (Quantitative Analysts) continuously iterate and improve on live algorithms, as we know that the world changes and our algorithms need to be up-to-date!

The success rate of this process is about 2-5% with lots of algorithms being rejected during this process. The algorithms that are currently live have already been running and improved on for at least five years. It may take months from ideation to the launch of a new algorithm!

Links to related articles:
The Algorithms
Algorithms at Unhedged

Information contained here is prepared by Unhedged Pty Ltd ACN 646 421 730 (AR 001288751) an authorised representative of Cache Investment Management Pty Ltd ACN 624 306 430 AFSL 514 360. Units in the Unhedged Fund are issued by Melbourne Securities Corporation Limited (ACN 160 326 545 AFSL 428 289). All information is general information only and doesn’t take into consideration your personal circumstances, financial situation or needs. Before making a financial decision you should understand the risks and read the relevant PDS and TMD and consider whether the product is right for you and seek advice if necessary. Any information contained here may have been automatically generated by an algorithm based on raw data inputs and has not been independently verified and is subject to change. Past performance is not indicative of future performance. Unhedged Fund APIR Code: UPL9838AU International Securities Identification Number (ISIN): AU60UPL98384. Copyright (C) 2022 Unhedged. All rights reserved.
Was this article helpful?
Thank you!