[ T H O U G H T
L E A D E R S H I P
|
L I Q U I D M E T R I X ]
Machine learning
engineering for TCA
Chris Sparrow, head of research, and Melinda Bui, director of trading analytics, at
LiquidMetrix explain how applying domain knowledge to its database of client
orders can create features that help get the most out of the machine learning algos.
W
hen we analyse our
trading performance
using transaction cost
analysis (TCA), we need to realise
there are things that are not in our
control, like the liquidity available
in the market, and things that are
in our control, like how we inter-
act with the market - our trading
strategy. The trading strategy
describes the prices and volumes
we obtained in the market and
include the venue distribution and
the time distribution of volume
executed passively and actively, lit
and dark, etc.
The point of TCA is to uncover
insights in our execution data.
These insights can shed light on
how to best model market impact
of our orders along with many
other use cases. Other use cases
include comparing how different
algos trade different securities,
how stable the trading strategies
are and many others.
We can use machine learning
(ML) algorithms to help us extract
26 // TheTrade // Summer 2019
these insights from our historical
set of order data. In this article we
explain a key part of the process of
applying ML to TCA - applying our
domain knowledge to our database
of client orders to create features
that help us get the most out of the
ML algos.
The trading strategy is often dic-
tated by the type of algo employed,
along with specific parameters
that modify the default strategy.
Depending on our reasons for
trading and the liquidity charac-
teristics of the stock, we may want
to modify our trading strategy.
Since we measure the perfor-
mance of each order using various
benchmarks, we can analyse how
performance varies with trading
strategy.
When we talk about a trading
strategy, we generally mean the
volume profile of our order as a
function of time. There are other
components related to venue
selection (eg: whether dark pools
are accessed, etc…) which can also
be measured to understand the
way our order interacted with the
market. We can then take these
measurements and create ana-
lytics, or features, to describe the
trading strategy.
We are particularly interested in
the trading strategy because that is
how we interacted with the market
and so we expect that for our pri-
mary use case, measuring market
impact, the degree of impact will
depend in part on our trading
strategy.
To put this another way, we want
to be able to do scenario analysis
using our pre-trade market impact
model, and we want to see how