Applications of Computational Intelligence in Data-Driven Trading by Cris Doloc

Author:Cris Doloc
Language: eng
Format: epub
ISBN: 9781119550518
Publisher: Wiley
Published: 2019-10-29T00:00:00+00:00

7.1.2 New Challenges

The primary challenge for the financial practitioners is not necessarily technological in nature, but it mostly relates to the implementation process. If a computer scientist strives to get extremely high accuracy for their ML models (usually > 99%), a quantitative financial practitioner considers anything above 50% accuracy a success. The field of Quantitative Finance is not a Silicon Valley type culture where open-source code and collaboration is praised or even encouraged. Quantitative Finance has historically thrived on proprietary models and trade secrets. Therefore, according to Guida, “One cannot do a pure copy-paste from computer science into finance.”

Again the challenge of the practical implementation is going to supersede the technological aspect. The primary focus of applying ML techniques in Quantitative Finance is to derive computationally and informationally efficient algorithms for inferring good predictive models from large data sets. A natural candidate for application of ML is to problems arising in High-Frequency Trading, for both trade execution and alpha generation.

One major challenge to overcome in training ML models on financial data is coping with the market dynamics, or the market microstructure. Understanding and modeling the dynamics of the price-generating processes is a central aspect for the applicability of ML techniques in Quantitative Finance. It is just unpractical to train an algorithm to categorize and classify all the permutations of every potential market dynamics scenario. From this perspective Market Microstructure is a very low signal-to-noise problem domain, sometimes calling into question the applicability of ML techniques.

While classic quantitative financial models usually prescribe what the relevant features for predictive modeling should be (i.e. excess returns, book-to-market ratios) in many HFT problems one may not have much prior intuition about what the relevant features should be. One typical question is how the distribution of liquidity in the order book is relating to future price movements, if at all. As such the process of selecting the relevant modeling features (feature engineering) is becoming the central theme for the use of ML in HFT.

Artificial Neural Networks are considered to be one of the most emblematic and utilized Machine Learning techniques nowadays. ANNs are considered to be universal function approximators and as such they are extremely flexible and may produce highly nonlinear functions of arbitrary and essentially uncontrollable complexity (e.g. highly non-smooth functions). As a consequence of being universal function approximators, the ANNs exhibit both notable strengths and weaknesses. Let's consider the example of using a generic ANN to build a financial forecasting model. Such a model will most likely be a highly nonlinear function of the input variables, but of an unknown form. It could be as simple as a cubic polynomial or it could take a much more complex mathematical form.

In ML parlance, if the model passes an out-of-sample test, it is considered to be acceptable no matter the functional form it may take. But when one attempts to apply this model to any market data set, one may be surprised by the results. If the market regime has changed since the time when

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.