In the last course of our specialization, Overview of Advanced Methods of Reinforcement Learning in Finance, we will take a deeper look into topics discussed in our third course, Reinforcement Learning in Finance. In particular, we will talk about links between Reinforcement Learning, option pricing and physics, implications of Inverse Reinforcement Learning for modeling market impact and price dynamics, and perception-action cycles in Reinforcement Learning. Finally, we will overview trending and potential applications of Reinforcement Learning for high frequency trading, cryptocurrencies, peer-to-peer lending, and more.
LOB Modeling
Loading...
From the course by New York University Tandon School of Engineering
Overview of Advanced Methods of Reinforcement Learning in Finance
15 ratings
New York University Tandon School of Engineering
15 ratings
Course 4 of 4 in the Specialization Machine Learning and Reinforcement Learning in Finance
From the lesson
Other Applications of Reinforcement Learning: P-2-P Lending, Cryptocurrency, etc.
Meet the Instructors
Igor Halperin
Now, after we talked about what the LOB is,
let's talk about how we can model it.
As a short summary of what we discussed so far,
an LOB aggregates both sides of the market.
It gives multi-level summary of the state of supply and demand sides of the market,
and there is a substantial empirical research that shows that the state
of the LOB is very informative of a short-time future.
This means that the short-time price movements are quite
predictable if we condition our predictions on the state of the LOB.
In other words, we can say that features constructed of
the LOB data should be useful for building short-term forecast models.
Understanding of the mechanics of the LOB is
quite important for a number of applications.
One of them is optimization of trade execution and minimization of marketing,
but we already mentioned
both these topics in our previous course on reinforcement learning.
Other applications deal with the design optimal in today trading strategies.
Understanding of the LOB is important even
for those players that operate that different frequencies,
for example, daily frequency.
The reason for this is a propagation of dynamics
and signals between different timescales in the market.
One such example of connections between high-frequency trading can
traditional low-frequency daily trading is the Flash Crash of 2010.
You can read about this event in the agent list for this week.
Now, let's talk a bit about the differences between
a portfolio allocation problem and an optimal placement problem for the LOB.
First, for portfolio allocation,
the event horizon is in weeks or months or even years.
Traits for this setting usually extend from minutes to days,
and portfolio managers solve their own problems of portfolio optimization.
The cost of execution for them is a fixed cost that they pay to brokers.
Now, for brokers, the task is to keep the largest part of the sphere as a profit.
So, they have to solve their own optimization problem,
namely the problem is now how to split the market order received from the trader
optimally and this translates into a stream of
orders that they start to place at very high frequency.
Their decision space may be focused on
only one stock but it's still pretty high dimensional space.
The broker should decide on the time sequence of
orders and for each order should be either a market or limit order.
It should also specify the order size and it should specify the venue,
namely to what exchange the order should be submitted.
So, together, this produces a high-dimensional decision space even for a single stock.
So, the decision space is high-dimensional and discrete.
If we want to trade a few stocks simultaneously,
the dimensionality will increase exponentially.
The dynamics of the LOB is also quite complex as it has so many moving parts.
So, a big question is,
"What kind of models will work best for such settings?"
There is one view that says that
simple and analytically attractable models should be preferred.
In particular, one reason for this is that those models should work fast.
Another view is that the dynamics are so complex here that
we better should rely on machine learning methods that are model free.
A bit more general view of different approaches to modeling
the LOB is to classify them into three categories.
The first class is formed by economics-based equilibrium approaches.
In this approach is a single agent is involved in
the sequential game against the market represented by a state of the LOB.
Other approaches focus directly on modal and observables of the LOB.
For example, statistical and machine learning methods
focus on statistical patterns of order flows,
and they do not try to explain the mechanics of the LOB.
Instead, they operate with exogenous random processes they describe order flows.
Finally, there are physics motivated approaches that have
you read them order flow says connected to
stochastic dynamics of interactions within the LOB.
We have already mentioned many potential applications of
models of the LOB and here I list them again for you.
They include market-making, optimal execution,
risk management, daily strategies,
option trading, and regulation.
Finally, I want to talk about different types of data relevant for the LOB.
So, depending on their level of aggregation,
there are three levels of
the LOB data which is something that we already mentioned before.
So, to reiterate, Level-1 data,
also called the quotes data,
shows only the top of the LOB on both sides.
In other words, it only shows the best bid and ask prices and the total liquidity,
that is the number of trades available at this level,
and this is the most basic information from the LOB that can be
combined with the trading data as we'll discuss a bit later.
The second level is called Level-2 data and it shows the aggregated state
of the LOB for five levels into both the demand and supply sides.
So, it gives you five bid quotes with quantities and five ask quotes with quantities.
Finally, Level-3 data is the most defined data that
provides information about each individual limit order.
Out of these categories,
Level-1 data is the most accessible and cheapest.
There are a number of providers of such data, for example, tickdata.com.
Level-2 data is considerably more expensive
because it's the most interesting available data for market participants.
It provides considerably more information than
Level-1 data and many market players prefer to use this data.
Finally, Level-3 data is of course the most valuable piece because it provides
complete data about these data are generally available only to regulators.
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2018 Coursera Inc. All rights reserved.
