0:09

And then last week, we talked about using

the reinforcement learning for option pricing and hedging.

We built this simple and analytically tractable reinforcement learning model

that solves the most fundamental problem of option pricing,

the problem of pricing and hedging over a single European option,

which was a put option in our case,

on a single stock.

Now, we want to see how reinforcement learning applies to stock trading.

You may wonder at this point,

why did we first talk about options on stocks before talking about stocks themselves.

The reason for this is that our first case is probably

a simplest possible reinforcement learning setting in finance,

that has practical interest on its own.

As we will see shortly,

applications of reinforcement learning to stock trading

are more technically involved than this example,

for a number of reasons.

And to see that,

it might be good to start talking about

applications of reinforcement learning for stock trading,

with a brief summary of what we did for options.

So, what we found was that,

by the classical option replication argument of Blokes,

Black, Scholes, and Netherton,

pricing of an option on the stock amounts to

dynamic optimization of a very simple portfolio made of stock and cash.

The objective insights problem is the best dynamic application of the option.

We then found that if we consider risk-adjusted return of

such portfolio as a one-step reward function for reinforcement learning formulation,

then we can apply the standard risk neutral formulation of Q-learning,

which only optimizes expected rewards,

but does not control variance of these rewards.

But this is okay in our case because we already

included the risk penalty in our risk-adjusted expected rewards.

Another important observation was that rewards in our formulations and

our formulation were quadratic only for non-zero risk-averse parameter Lambda.

Also important was the fact that, in this problem,

we took a view of a small investor,

whose trades in stock do not impact the market.

We also neglected transaction costs for simplicity, but in principle,

transaction costs can be introduced in our reinforcement learning option pricing model,

by making the model only a little bit more complex.

Yet, in the simplest setting without transaction cost that we considered,

the problem was easy to solve because of two related reasons.

The first reason is that,

when there are no transaction costs,

we do not need to store the number of stocks we currently hold as a state variable.

In fact, in our previous model,

it was an action variable instead.

The second reason is that,

we assumed a small investor, and therefore,

emitted market impact effects which would make the whole problem non-linear.

These two simplifications made it relatively easy to solve the problem of

option pricing and hedging using Q-learning and fitted Q iteration.

Now, let's see what changes when we start to look

at a related portfolio tasks for stock trading.

In general, most of practical problems in stock trading can be viewed

as problems of optimization of stock portfolios.

Let me first give you a list of such problems,

and then we will go in a little bit more details over each one of them in this list.

The first example in this list would be the problem of

optimal trade execution for a large trade in stock over a given company.

In this case, we have a stock portfolio made off the stock of the same company.

The second group of examples involves multi-asset portfolios made of different stocks.

This different but related tasks include;

portfolio dynamic management, optimal portfolio liquidation, and index tracking.

We will discuss shortly what is involved in each one of these tasks but for now,

let's first outline what is common between all of them.

First, as we will see later,

all of these problems can be formulated as problems with

a one one-step rewards made of an expected cost of trading plus risk penalties.

Second, unlike our option pricing case,

now we need to keep stock positions at state variables.

And third, in this class of problems,

we often have to take into account

a feedback loop for an impact of trading on stock prices.

Moreover, many of these problems are high dimensional,

even for a single stock,

and many of them are very high dimensional for both the state and actions spaces.

And finally, often, these problems have to do with Big Data,

in particular, for large portfolios or for high frequency trading.

We will talk more about the problems involving Big Data, deep reinforcement learning,

and related topics in our next course, but here,

I just want you to take a mental note of this fact.

It's important in particular because,

this very fact rules out some traditional modeling approaches,

such as dynamic problem, right from the start.

And the reason is that the classical dynamic programming is

based on a discretization of a state space,

and this only works when you have at most,

three or maybe four dimensional stage space.

But the dimensionality of state space in

all problems that we're going to discuss in this week is typically much higher,

in tens, hundreds, or even thousands.

So, for these cases,

a discrete state space is a non-starter.

And hence, we will not discuss such formulations here.

Now, let's go over each one of these portfolio tasks in more details.

First, let's talk about optimal stock execution.

The problem in this case is formulated as follows: We have

a large number M of stocks of a particular company, for example, Amazon,

and we want to sell all the stocks within some time,

T. In many practical cases, the time T,

which we will call the planning horizon, is in minutes.

This is so-called agency problem,

which is solved thousands of times a day by brokers on the stock market exchange,

following sell or buy orders by their clients, typically,

hedge funds or asset managers,

that we can collectively call traders.

Sometimes, traders may execute

trades themselves if they have an appropriate infrastructure.

What these brokers or traders do when they have to sell

a large block of a stock is they break it into smaller chunks,

and sell or buy these chunks sequentially.

So that to minimize markets impact from their trades.

Mathematically, this amounts to dynamic optimization with a feedback loop,

which appears as a result of market impact.

And additional complexity is brought to the problem by the fact

that there are two sorts of orders that the broker can receive from a client;

market order and limit order.

A market order is to sell or buy

a certain number of stocks at the prevailing market price.

A limit order is to sell or buy for a pre-specified price,

but without a guarantee that the trade will be made for the price requested.

If we take into account both market and limit orders,

which would be the right thing to do for high frequency trading,

the problem may quickly become a high dimensional and pretty complex,

even for a single stock.

We will talk more about such staging in our next course in this specialization.

Now, let's talk about the problem of optimal portfolio liquidation.

The problem is nearly the same as the last problem, except that now,

we start with a portfolio made of different stocks,

rather than with a block of stock of the same company.

Also because we now talk about a portfolio liquidation strategy,

it might work on a bit longer planning horizon,

for example, hours or days.

The next problem is dynamic portfolio optimization problem.

This is essentially the previous problem with the minus sign where,

instead of liquidating given starting portfolio,

would built an optimal investment portfolio.

An objective function for such dynamic optimization is

defined as a risk adjusted and cost adjusted portfolio return,

and is mathematically identical to the objective function of the previous problem.

The main difference is quantitative rather than qualitative.

It has to do with the fact that,

planning horizon for such tasks is longer and may extend to weeks,

months, or even years.

And therefore, a different set of predictors is needed for such longer planning horizons.

For example, we have to include macro-economic factors and other factors.

And please note that

the classical markets portfolio optimization problem

is a special one-step version of this setting.

These problems can also easily become

high dimensional and Big Data problems depending on a specified investment universe.

Finally, I want to talk about the problem of optimal index tracking.

Here, a task is typically to track a market index,

such as an S&P 500 portfolio,

using a proxy portfolio made of a smaller number of n different stocks,

which can be viewed as the main drivers of the index itself.

Such tasks can have plain horizons,

so week, months, or can have an infinite horizon.

The latter scenario of infinite horizon arises if you

invest in the so-called Exchange Traded Funds or ETFs for short.

The idea of these instruments was exactly this,

to construct a very liquid exchange traded portfolio over

a reasonably small number of stocks that would causally match by design markets indices.

Because indices such as S&P 500 are just indices,

you cannot directly invest in them.

But if you invest in a ETF called SPY,

you will get the best proxy portfolio for the index for a reasonable price.

For other markets indices,

or other given target client-specific portfolios,

we can also formulate the problem of optimal tracking.

And a mathematical formulation will be very similar to the previous one,

except for the terminal condition that we have to use.

So, here is a summary and a plan of what we will do next.

First, all these problems are problems of optimal control.

And therefore, we will consider reinforcement learning methods for such problems.

Second, all these problems are high dimensional problems.

Therefore, we will not discretize state or action space.

But we will work directly with a continuous formulation.

And third, we will present a simple,

but general mathematical formulation that works for all the problems we outlined above.

This approach will be quite tractable analytically

as long as it's based on convex optimization.

Let's see what questions we could ask ourselves after this introduction,

and then continue in the next video.