# 13 Option pricing via machine learning

Machine learning (ML) is seen as a part of artificial intelligence.
ML algorithms build a model based on training data in order to make predictions or decisions without being explicitly programmed to do so.
While ML can be specified along a vast array of different branches, this chapter focuses on so-called supervised learning for regressions. The basic idea of supervised learning algorithms is to build a mathematical model for data that contains both the inputs and the desired outputs. In this chapter, we apply well-known methods such as random forests and neural networks to a simple application in option pricing. More specifically, we are going to create an artificial dataset of option prices for different values based on the Black-Scholes pricing equation for call options. Then, we train different models to *learn* how to price call options without prior knowledge of the theoretical underpinnings of the famous option pricing equation by Black and Scholes (1973).

The roadmap of this chapter is as follows: First, we briefly introduce regression trees, random forests, and neural networks. As the focus is on implementation, we leave a thorough treatment of the statistical underpinnings to other textbooks from authors with a real comparative advantage on these issues.
We show how to implement random forests and deep neural networks with tidy principles using `tidymodels`

or `tensorflow`

for more complicated network structures.

In order to replicate the analysis regarding neural networks in this chapter, you have to install `TensorFlow`

on your system, which requires administrator rights on your machine. Parts of this can be done from within R. Just follow these quick-start instructions.

Throughout this chapter, we need the following packages.

The package `tidymodels`

(Kuhn and Wickham 2020) provides the tidyverse’s take on machine learning and is an opinionated collection of packages that provide a consistent workflow for common machine learning tasks. `keras`

(Allaire and Chollet 2022) is a high-level neural networks API developed with a focus on enabling fast experimentation with Tensorflow.
The package `ranger`

(Wright and Ziegler 2017) provides a fast implementation for random forests and `hardhat`

(Vaughan and Kuhn 2022) is a helper function to for robust data preprocessing at fit time and prediction time.

## 13.1 Regression trees and random forests

Regression trees are a popular ML approach for incorporating multiway predictor interactions. In Finance, regression trees are used increasingly, also in the context of asset pricing, see, e.g. Bryzgalova, Pelger, and Zhu (2022). Trees are fully nonparametric and possess a logic that departs markedly from traditional regressions. Trees are designed to find groups of observations that behave similarly to each other. A tree “grows” in a sequence of steps. At each step, a new “branch” sorts the data leftover from the preceding step into bins based on one of the predictor variables. This sequential branching slices the space of predictors into rectangular partitions and approximates the unknown function \(f(x)\) with the average value of the outcome variable within each partition. For a more thourough treatment of regression trees, we refer to Coqueret and Guida (2020).

Formally, partition the predictor space into \(J\) non-overlapping regions, \(R_1, R_2, \ldots, R_J\). For any predictor \(x\) that falls within region \(R_j\) we estimate \(f(x)\) with the average of the training observations, \(\hat y_i\), for which the associated predictor \(x_i\) is also in \(R_j\). Once we select a partition \(\mathbf{x}\) to split in order to create the new partitions, we find a predictor \(j\) and value \(s\) that define two new partitions, called \(R_1(j,s)\) and \(R_2(j,s)\), which split our observations in the current partition by asking if \(x_j\) is bigger than \(s\): \[R_1(j,s) = \{\mathbf{x} \mid x_j < s\} \mbox{ and } R_2(j,s) = \{\mathbf{x} \mid x_j \geq s\}.\] To pick \(j\) and \(s\), we find the pair that minimizes the residual sum of square (RSS): \[\sum_{i:\, x_i \in R_1(j,s)} (y_i - \hat{y}_{R_1})^2 + \sum_{i:\, x_i \in R_2(j,s)} (y_i - \hat{y}_{R_2})^2\] Note: Unlike the case for the sample variance, we do not scale by the number of elements \(R_k(j, s)\)! As in the chapter on penalized regressions, the first relevant question is: What are the hyperparameter decisions? Instead of a regularization parameter, trees are fully determined by the number of branches used to generate a partition (sometimes one specifies the minimum number of observations in each final branch instead of the maximum number of branches).

Models with a single tree suffer from high variance. Random forests address the shortcomings of decision trees. The goal is to improve the predictive performance and reduce instability by averaging multiple decision trees (a forest of trees constructed with randomness). A forest basically implies creating many regression trees and averaging their predictions. To assure that the individual trees are not the same, we use a bootstrap to induce randomness. More specifically, we build \(B\) decision trees \(T_1, \ldots, T_B\) using the training sample. For that purpose, we randomly select features to be included in the building of each tree. For each observation in the test set we then form a prediction \(\hat{y} = \frac{1}{B}\sum\limits_{i=1}^B\hat{y}_{T_i}\).

## 13.2 Neural networks

Roughly speaking, neural networks propagate information from an input layer, through one or multiple hidden layers, to an output layer. While the number of units (neurons) in the input layer is equal to the dimension of the predictors, the output layer usually consists of one neuron (for regression) or multiple neurons for classification. The output layer predicts the future data, similar to the fitted value in a regression analysis. Neural networks have theoretical underpinnings as “universal approximators” for any smooth predictive association (Hornik 1991). Their complexity, however, ranks neural networks among the least transparent, least interpretable, and most highly parameterized ML tools. In Finance, applications of neural networks can be found in the context of many different applications, e.g. Avramov, Cheng, and Metzker (2022), Gu, Kelly, and Xiu (2020) and L. Chen, Pelger, and Zhu (2019).

Each neuron applies a nonlinear “activation function” \(f\) to its aggregated signal before sending its output to the next layer \[x_k^l = f\left(\theta^k_{0} + \sum\limits_{j = 1}^{N ^l}z_j\theta_{l,j}^k\right)\] While the easiest case with \(f(x) = \alpha + \beta x\) resembles linear regression, typical activation functions are sigmoid (i.e., \(f(x) = (1+e^{-x})^{-1}\)) or ReLu (i.e., \(f(x) = max(x, 0)\)).

Neural networks gain their flexibility from chaining multiple layers together. Naturally, this imposes many degrees of freedom on the network architecture for which no clear theoretical guidance exists. The specification of a neural network requires, at a minimum, a stance on depth (number of hidden layers), the activation function, the number of neurons, the connection structure of the units (dense or sparse), and the application of regularization techniques to avoid overfitting. Finally, *learning* means to choose optimal parameters relying on numerical optimization, which often requires specifying an appropriate learning rate.

Despite the computational challenges, implementation in R is not tedious at all because we can use the API to `tensorflow`

.

## 13.3 Option pricing

To apply ML methods in a relevant field of finance, we focus on option pricing. The application in its core is taken from Hull (2020). In its most basic form, call options give the owner the right but not the obligation to buy a specific stock (the underlying) at a specific price (the strike price \(K\)) at a specific date (the exercise date \(T\)). The Black–Scholes price (Black and Scholes 1973) of a call option for a non-dividend-paying underlying stock is given by \[ \begin{aligned} C(S, T) &= \Phi(d_1)S - \Phi(d_1 - \sigma\sqrt{T})Ke^{-r T} \\ d_1 &= \frac{1}{\sigma\sqrt{T}}\left(\ln\left(\frac{S}{K}\right) + \left(r_f + \frac{\sigma^2}{2}\right)T\right) \end{aligned} \] where \(C(S, T)\) is the price of the option as a function of today’s stock price of the underlying, \(S\), with time to maturity \(T\), \(r_f\) is the risk-free interest rate, and \(\sigma\) is the volatility of the underlying stock return. \(\Phi\) is the cumulative distribution function of a standard normal random variable.

The Black-Scholes equation provides a way to compute the arbitrage-free price of a call option once the parameters \(S, K, r_f, T\), and \(\sigma\) are specified (arguably, all parameters are easy to specify except for \(\sigma\) which has to be estimated). A simple R function allows computing the price as we do below.

## 13.4 Learning Black-Scholes

We illustrate the concept of ML by showing how ML methods *learn* the Black-Scholes equation after observing some different specifications and corresponding prices without us revealing the exact pricing equation.

### 13.4.1 Data simulation

To that end, we start with simulated data. We compute option prices for call options for a grid of different combinations of times to maturity (`T`

), risk-free rates (`r`

), volatilities (`sigma`

), strike prices (`K`

), and current stock prices (`S`

). In the code below, we add an idiosyncratic error term to each observation such that the prices considered do not exactly reflect the values implied by the Black-Scholes equation.

```
option_prices <- expand_grid(
S = 40:60, # Stock price
K = 20:90, # Strike price
r = seq(from = 0, to = 0.05, by = 0.01), # Risk-free rate
T = seq(from = 3 / 12, to = 2, by = 1 / 12), # Time to maturity
sigma = seq(from = 0.1, to = 0.8, by = 0.1) # Volatility
) |>
mutate(
black_scholes = black_scholes_price(S, K, r, T, sigma),
observed_price = map(
black_scholes,
function(x) x + rnorm(2, sd = 0.15)
)
) |>
unnest(observed_price)
```

The code above generates more than 1.5 million random parameter constellations. For each of these values, two *observed* prices reflecting the Black-Scholes prices are given and a random innovation term *pollutes* the observed prices.

Next, we split the data into a training set (which contains 1% of all the observed option prices) and a test set that will only be used for the final evaluation. Note that the entire grid of possible combinations contains 3148992 different specifications. Thus, the sample to learn the Black-Scholes price contains only 31489 observations and is therefore relatively small.
In order to keep the analysis reproducible, we use `set.seed()`

. A random seed specifies the start point when a computer generates a random number sequence and ensures that our simulated data is the same across different machines.

```
set.seed(420)
split <- initial_split(option_prices, prop = 1 / 100)
```

We process the training dataset further before we fit the different ML models. We define a `recipe`

that defines all processing steps for that purpose. For our specific case, we want to explain the observed price by the five variables that enter the Black-Scholes equation. The *true* price (stored in column `black_scholes`

) should obviously not be used to fit the model. The recipe also reflects that we standardize all predictors to ensure that each variable exhibits a sample average of zero and a sample standard deviation of one.

```
rec <- recipe(observed_price ~ .,
data = option_prices
) |>
step_rm(black_scholes) |>
step_normalize(all_predictors())
```

### 13.4.2 Single layer networks and random forests

Next, we show how to fit a neural network to the data. Note that this requires that `keras`

is installed on your local machine. The function `mlp()`

from the package `parsnip`

provides the functionality to initialize a single layer, feed-forward neural network. The specification below defines a single layer feed-forward neural network with 10 hidden units. We set the number of training iterations to `epochs = 500`

. The option `set_mode("regression")`

specifies a linear activation function for the output layer.

```
nnet_model <- mlp(
epochs = 500,
hidden_units = 10,
) |>
set_mode("regression") |>
set_engine("keras", verbose = FALSE)
```

The `verbose=0`

argument prevents logging the results. We can follow the straightforward `tidymodel`

workflow as in the chapter before: Define a workflow, equip it with the recipe, and specify the associated model. Finally, fit the model with the training data.

```
nn_fit <- workflow() |>
add_recipe(rec) |>
add_model(nnet_model) |>
fit(data = training(split))
```

Once you are familiar with the `tidymodel`

workflow, it is a piece of cake to fit other models from the `parsnip`

family.
For instance, the model below initializes a random forest with 50 trees contained in the ensemble where we require at least 2000 observations in a node.
The random forests are trained using the package `ranger`

, which is required to be installed in order to run the code below (loading the package as we do above is not necessary, however, provides a transparent overview of the required packages for the session).

```
rf_model <- rand_forest(
trees = 50,
min_n = 2000
) |>
set_engine("ranger") |>
set_mode("regression")
```

Fitting the model follows exactly the same convention as for the neural network before.

```
rf_fit <- workflow() |>
add_recipe(rec) |>
add_model(rf_model) |>
fit(data = training(split))
```

### 13.4.3 Deep neural networks

Note that while the `tidymodels`

workflow is extremely convenient, more sophisticated *deep* neural networks are not supported yet (as of June 2022). For that reason, the code snippet below illustrates how to initialize a sequential model with three hidden layers with 10 units per layer. The `keras`

package provides a convenient interface and is flexible enough to handle different activation functions. The `compile()`

command defines the loss function with which the model predictions are evaluated.

```
model <- keras_model_sequential() |>
layer_dense(
input_shape = 5,
units = 10,
activation = "sigmoid"
) |>
layer_dense(units = 10, activation = "sigmoid") |>
layer_dense(units = 10, activation = "sigmoid") |>
layer_dense(units = 1, activation = "linear") |>
compile(
loss = "mean_squared_error"
)
model
```

```
Model: "sequential_1"
_________________________________________________________________________________
Layer (type) Output Shape Param #
=================================================================================
dense_5 (Dense) (None, 10) 60
dense_4 (Dense) (None, 10) 110
dense_3 (Dense) (None, 10) 110
dense_2 (Dense) (None, 1) 11
=================================================================================
Total params: 291
Trainable params: 291
Non-trainable params: 0
_________________________________________________________________________________
```

To train the neural network, we provide the inputs (`x`

) and the variable to predict (`y`

) and then fit the parameters. Note the slightly tedious use of the method `extract_mold(nn_fit)`

. Instead of simply using the *raw* data, we fit the neural network with the same processed data that is used for the single-layer feed-forward network. What is the difference to simply calling `x = training(data) |> select(-observed_price, -black_scholes)`

? Recall that the recipe standardizes the variables such that all columns have unit standard deviation and zero mean. Further, it adds consistency if we ensure that all models are trained using the same recipe such that a change in the recipe is reflected in the performance of any model. A final note on a potentially irritating observation: Note that `fit()`

alters the `keras`

model - this is one of the few instances where a function in R alters the *input* such that after the function call the object `model`

is not same anymore!

```
model |>
fit(
x = extract_mold(nn_fit)$predictors |> as.matrix(),
y = extract_mold(nn_fit)$outcomes |> pull(observed_price),
epochs = 500, verbose = FALSE
)
```

### 13.4.4 Universal approximation

Before we evaluate the results, we implement one more model: In principle, any non-linear function can also be approximated by a linear model containing the input variables’ polynomial expansions. To illustrate this, we first define a new recipe, `rec_linear`

, which processes the training data even further. We include polynomials up to the fifth degree of each predictor and then add all possible pairwise interaction terms. The final recipe step, `step_lincomb()`

, removes potentially redundant variables (for instance, the interaction between \(r^2\) and \(r^3\) is the same as the term \(r^5\)). We fit a Lasso regression model with a pre-specified penalty term (consult the chapter on factor selection on how to tune the model hyperparameters).

```
rec_linear <- rec |>
step_poly(all_predictors(),
degree = 5,
options = list(raw = T)
) |>
step_interact(terms = ~ all_predictors():all_predictors()) |>
step_lincomb(all_predictors())
lm_model <- linear_reg(penalty = 0.01) |>
set_engine("glmnet")
lm_fit <- workflow() |>
add_recipe(rec_linear) |>
add_model(lm_model) |>
fit(data = training(split))
```

## 13.5 Prediction evaluation

Finally, we collect all predictions to compare the *out-of-sample* prediction error evaluated on ten thousand new data points. Note that for the evaluation, we use the call to `extract_mold()`

to ensure that we use the same pre-processing steps for the testing data across each model. We also use the somewhat advanced functionality in `forge()`

, which provides an easy, consistent, and robust pre-processor at prediction time.

```
out_of_sample_data <- testing(split) |>
slice_sample(n = 10000)
predictive_performance <- model |>
predict(forge(
out_of_sample_data,
extract_mold(nn_fit)$blueprint
)$predictors |> as.matrix()) |>
as.vector() |>
tibble("Deep NN" = _) |>
bind_cols(nn_fit |>
predict(out_of_sample_data)) |>
rename("Single layer" = .pred) |>
bind_cols(lm_fit |> predict(out_of_sample_data)) |>
rename("Lasso" = .pred) |>
bind_cols(rf_fit |> predict(out_of_sample_data)) |>
rename("Random forest" = .pred) |>
bind_cols(out_of_sample_data) |>
pivot_longer("Deep NN":"Random forest", names_to = "Model") |>
mutate(
moneyness = (S - K),
pricing_error = abs(value - black_scholes)
)
```

In the lines above, we use each of the fitted models to generate predictions for the entire test data set of option prices. We evaluate the absolute pricing error as one possible measure of pricing accuracy, defined as the absolute value of the difference between predicted option price and the theoretical correct option price from the Black-Scholes model.

```
predictive_performance |>
ggplot(aes(x = moneyness, y = pricing_error, color = Model)) +
geom_jitter(alpha = 0.05) +
geom_smooth(se = FALSE, method = "gam") +
labs(
x = "Moneyness (S - K)", color = NULL,
y = "Absolut prediction error (USD)",
title = "Prediction errors: Call option prices"
)
```

The results can be summarized as follow: i) All ML methods seem to be able to *price* call options after observing the training test set. ii) The average prediction errors increase for far in-the money options. ii) Random forest and the Lasso seem to perform consistently worse in prediction option prices than the Neural networks. iii) The complexity of the deep neural network relative to the single layer neural network does not result in better out-of-sample predictions.

## 13.6 Exercises

- Write a function that takes
`y`

and a matrix of predictors`X`

as inputs and returns a characterization of the relevant parameters of a regression tree with**1**branch. - Create a function that creates predictions for a new matrix of predictors
`newX`

based on the estimated regression tree. - Use the package
`rpart`

to*grow*a tree based on the training data and use the illustration tools in`rpart`

to understand which characteristics the tree deems relevant for option pricing. - Make use of a training and a test set to choose the optimal depth (number of sample splits) of the tree.
- Use ‘keras’ to initialize a sequential neural network that can take the predictors from the training dataset as input, contains at least one hidden layer, and generates continuous predictions.
*This sounds harder than it is:*see a simple regression example here. How many parameters does the neural network you aim to fit have? - Next, compile the object. It is important that you specify a loss function. Illustrate the difference in predictive accuracy for different architecture choices.