Some fifteen years after Nitin Gambhir founded financial software firm Tethys Technology, its technology arm has emerged as a multi-award winning provider of trading algorithms. For the past decade, the Tethys execution research team has been working on machine learning (ML) and other artificial intelligence (AI) techniques, and incorporating these in its algorithms.
Gambhir, the chief executive of Tethys Technology, the leading independent provider of execution algorithms for global equities, futures, options and FX, is an expert in the application of AI to trading markets but admits the industry still has a long way to go in its use of these methods.
“At a high level, financial markets are ideal for machine learning and AI techniques. There is a lot of data, a lot of factors, and the inter-relationships between factors are complex. Machine learning embodies non-linear techniques that facilitate the discovery of relationships that would otherwise be difficult to recognise with classical statistical techniques. Despite this, failures in applying ML/AI to trading and alpha models are not uncommon. Overfitting, and resulting poor out-of sample performance, is the norm.”
Gambhir said financial markets are dynamic and relationships therein are in flux at all times. “There are unlimited potential explanatory factors and signal to noise ratios are very low. Relationships appear and go away. It is hard to discern if this is because of regime sensitivity or observed relationship was just a random occurrence. Stable out of sample relationship identification is challenging. The industry is rife with examples of capable statisticians and AI experts failing to achieve success.”
The Tethys chief argues the main reasons for these failings are the “lack of understanding of the financial markets and thus, the failure to recognise when no real economic drivers hold a model together” and the "lack of a proper portfolio approach”.
For his part, Gambhir holds a BS from the Indian Institute of Technology in New Delhi, India, and a Masters from Yale University. He worked for five years at JP Morgan before joining Bear Stearns where he quickly rose to Managing Director. Thereafter, he worked at Point72 Asset Management before launching Tethys.
He continued: “At Tethys, we always start with an economic or behavioural hypothesis before formulating a model. We also believe that a model should be stable across market regimes and do well across asset classes and geography.”
Gambhir also challenges the popular assertion that machine learning is only as good as the underlying data: “We say it is only as good as the data engineering framework and the financial engineers who work on it. If you look at a largely stable data set, such as the factors that determined the efficiency of a cement factory for example, you can use AI to simply establish what you need to do to produce cement more efficiently.
“Financial markets change constantly. Models created on one set of data may not be effective on a different day when the market conditions are very different. In addition to only using sensible explanatory factors with limited co-linearity, we also use machine learning to do independent predictions of relevant factors. Our entire approach is that of bounded machine learning. If the models receive inputs that are not recognised (outside of expected bounds), the algorithms update in real-time and shift to over-weighting the alternatives, including more classical models.”
Gambhir said that his team uses machine learning and other AI techniques in reference to specific challenges. “Tethys does not espouse an all-encompassing ML execution algorithm, rather we use ML to tackle micro-tasks within the algorithm logic. We are not trying to achieve the equivalent of inventing a car that will drive itself under any condition. Instead, we firmly believe that finance requires a blended approach. We are using ML to make micro-predictions such as projected volumes over the next five minutes. A notable Tethys AI achievement is our success in event modelling. The anti-gaming components of our algorithmic suite leverage this extensively. Our AlgoWheel technology is another notable success.”
Gambhir said his firm will continue to use ML and AI to develop better and better trading algorithms. “Looking ahead, we are planning to deploy additional and larger data sets to look at a wider set of relationships. We are developing our own ML techniques that are faster and produce more stable results,” he concluded.