FX’s OTC nature will limit machine learning

Paul Golden
Published on:

Almost two decades after the principles of machine learning were applied to FX trading, challenges remain to be addressed for the technology to become ubiquitous.


A machine’s capacity to analyze market data in a quantity and speed that no human trader could hope to match is undeniable.

However, that machine also has to deal with data characteristics that are constantly moving, autocorrelation – where today’s price depends heavily on yesterday’s price – and low signal to noise ratios, which require it to sift through large volumes of meaningless data to find a meaningful result.


Kris Longmore,
Quantify Partners

The over-the-counter (OTC) nature of the FX market also means that data might only be applicable to specific brokers, says Kris Longmore, co-founder and head of quantitative research at Quantify Partners.

“Ideally, any forecasting algorithm would be robust to small differences in broker feeds, but still it is an issue,” he says. “Finally, the 24-hour nature of the market means that sampling time of market data will be a consideration.”

Celent senior analyst Joséphine de Chazournes refers to problems around latency, saying: “FX markets are a high-frequency asset class and machine learning does not support that kind of speed yet.

DeChazournes_Josephine 160x186
Joséphine de
Chazournes, Celent

“Strategies that work right now are long-short or event driven that learn overnight or at T+2 but never in real time, as putting a machine-learning algorithm within a data centre would be too costly.”

According to Peter Farley, senior strategist, capital markets, at Misys, firms will continue to invest in machine learning in the belief it will be inherently better than human trading. However, he reckons the fragmented nature of the FX market will discourage widespread adoption.

Machine learning will improve systematic models, but primarily as a more sophisticated dynamic error-correction methodology rather than a solution to the complex task of marrying long-term memory and experience with short-term weighting of new variables.

That is the view of Mark Farrington, managing director, portfolio manager and head of the macro currency group at Principal Global Investors, who suggests there are too many rotating long-term cycle analogues and dynamically changing short term variables to accurately assess, regress, apply weights and probabilities and then commit risk capital with conviction.

“Tactically positioning large portfolios quarter by quarter, delivering a stable annual absolute return for investors in alternatives, does not look to be the domain of machines,” he says.


Mark Farrington,
Principal Global

“If machine learning ever succeeds in mastering this segment of the market, it is likely to be due to the inclusion of one or two very smart human fund managers in the decision loop.”

From an information theoretical perspective, the influence of randomness on a time-based forecast target increases with time from the moment the prediction is made, so it would be reasonable to assume that short-term effects should be more predictable.

Forex Artilect founder David Lopez Onate suggests that high levels of noise mean it is preferable to create machine-learning models using longer-term data while using short-term charts as a trade trigger.

“That way you are trading with the signal of longer time frames, but in shorter time spans,” he says. “I believe machine learning is more effective for short-term strategies or scalping [making large numbers of trades for a small profit on each trade].”

In the short term, FX markets are mean reverting, whereas in the longer term they are trending, so different strategies are required for each, explains Yaron Golgher, CEO and co-founder of I Know First.

“Long-term traders cannot neglect data providing insight on forthcoming macroeconomic trends,” he says. “In the short term, it is more about news streams and trader behaviour that has to be learned from the data with greater granularity.”

Johannes Tynes,

Johannes Tynes, head of R&D at Inpirical, agrees that machine learning can be applicable to both short-term FX trading and longer-term strategies.

He notes that the data used in short-term trading encompasses non-financial and non-quantitative features – such as metrics derived from the natural language processing of live news streams and twitter patterns – to a greater extent, while the longer-term models tend to be more built around conventional macro variables and hypotheses.

“The model types also differ,” he concludes. “For the short term, classification models have more applications, although regression models [predicting exchange rate levels and quantities] are equally relevant for both.

“For short-term trading, online learning is often a must – continuous streams of data being fed into the model for updating and retraining – but for many longer-term applications it is best to do the model training, feature development, parameter tuning and validation on static data sets before deploying the model.”