Machine Learning Application to

Delta-Gamma Hedging
Machine Learning in Finance

Written at EPFL June 9, 2024

Group Members: Supervised by:

Chady Bensaid Elise Marie Gourier
Dorah Borgi
Simon Jonängen
William Olsson Teaching Assistant:
Sara Tegström Giuseppe Matera
 Abstract
Machine learning (ML) is rapidly transforming the financial industry by enabling sophisti-
cated data analysis, predictive modelling and automation of complex tasks for investors.
This projects aims to achieve a predictive model for creating an option portfolio that
is Delta-Gamma hedged towards risk. By using the equation for Black-Scholes option
pricing model for European options, different measurements of options such as Delta (∆)
and Gamma (Γ), commonly referred to as the Greeks, have been derived. Using these
expressions for the Greeks, regressive machine learning algorithms have then used in order
to continuously predict future values of Gamma. The main goal was to achieve a Delta-
Gamma neutral portfolio. Four different algorithms were used, namely, linear regression,
random forest, support vector regression and XGBoost. The algorithms were evaluated
under the two performance metrics of R2 and mean squared error (MSE).
Contents
1 Options as a Financial Derivative 1
1.1 Background information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Call Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Put Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Greeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2.1 Delta (∆) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Gamma (Γ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.3 Theta (Θ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Theory 3
2.1 Black-Scholes Option Pricing Model . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Derivation of Black-Scholes Price . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 The Greeks for the Black-Scholes . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.1 Delta (∆) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.2 Gamma (Γ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
∂Γ
2.3.3 Derivative of Gamma ( ∂S t
) . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.4 Theta (Θ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Hedging Strategy & Algorithmic Application 8

3.1 Delta-Neutrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Gamma-Neutrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 Delta-gamma hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.4 Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4.1 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4.2 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4.3 Gradient Boosting Machines . . . . . . . . . . . . . . . . . . . . . . . . 10
3.4.4 Support Vector Regression . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Data Extraction, Cleaning and Exploration 11

4.1 Data Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Stock Prices, Options Prices and Greeks . . . . . . . . . . . . . . . . . . . . . . 13

5 Implementation 18
5.1 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2 Delta-Gamma Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

6 Alternative Methods Results 21

6.1 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
6.2 SVR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6.3 XGBoost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

7 Methods comparison 22

8 Conclusion 23

References 24

A Appendix: Alternative approach i

1 Options as a Financial Derivative
Options are financial derivatives that give the buyer the right, but not the obligation, to buy or
sell an underlying asset at a specified price within a fixed period. They are used for hedging
risk or speculating on the price movements of assets like stocks, commodities, or currencies.

1.1 Background information

Options are financial derivatives that grant the buyer the right, but not the obligation, to buy
or sell an underlying asset at a specified price, known as the strike price, on or before a specified
date. These instruments are utilized for hedging risk, speculating on future price movements,
and generating income, across various underlying assets such as stocks, indices, commodities,
and currencies [1] .

Options are versatile tools in the financial markets, allowing for tailored investment strategies.
They are contracts with specified key details such as the underlying asset, strike price, expiration
date, and the premium. The value of an option is derived from the intrinsic value and the time
value, which decreases as the option nears expiration. The use of options spans across:
– Hedging: To protect against adverse price movements.
– Income Generation: Through premium collection from selling options.

– Speculation: Betting on the direction of the market with leveraged potential profits or
losses.
Options require a comprehensive understanding of market mechanics and the associated risks,
including the potential loss of the entire premium paid.

1.1.1 Call Options

A call option gives the holder the right to buy the underlying asset at the strike price within
a specified time period. Investors buy call options if they believe the price of the underlying
asset will rise above the strike price before the expiration date. If this happens, the investor can
purchase the asset at the lower strike price and potentially sell it at a higher market price, thus
realizing a profit [1].

1.1.2 Put Options

Put options give the holder the right to sell the underlying asset at the strike price within a
specified time period. Investors buy put options if they believe the price of the underlying asset
will fall below the strike price before the expiration date. This allows the investor to sell the
asset at the higher strike price, even if the market price has fallen, thus providing a mechanism
for hedging against potential losses or for speculative gains [1].

1.2 The Greeks

The Greeks are quantitative measures that describe how sensitive the price of an option is to
factors such as time decay, volatility, and changes in the value of the underlying asset.

1
1.2.1 Delta (∆)
In the context of call options, ”delta” (∆) is a measure of how much the price of an option is
expected to move based on a change of one unit of price of the underlying asset, such as one
dollar. It is one of the ”Greeks,” which are metrics used to assess the risk and potential reward
of options positions. Delta is particularly important because it gives investors and traders an
idea of how the value of an option might change as the market price of the underlying asset
moves, thereby helping in decision-making processes [2].

For call options, delta can range in the interval [0, 1]. A delta of 0 means the option’s price is
not expected to move in response to price changes in the underlying asset. A delta of 1 means
the option’s price is expected to move one-for-one with the price of the underlying asset, thus
being equivalent to holding a share of a stock.

Delta can also be interpreted as the option’s sensitivity to price changes in the underlying asset
or as an estimate of the probability that the option will expire in-the-money (ITM). A higher
delta not only indicates a greater sensitivity to changes in the underlying asset’s price but also
suggests a higher probability of the option expiring ITM. For instance, a delta of 0.75 suggests a
75% theoretical probability of expiring ITM, making it a more attractive choice for those bullish
on the underlying asset.

Options that are at-the-money (ATM) generally have a delta around 0.5, reflecting that they
have an approximately equal chance of ending up in or out of the money. As the underlying
asset’s price moves in the money for a call option, its delta will increase, approaching 1, indicating
it is moving deeper ITM and its price is more closely tracking the underlying asset.

Delta is not static; it changes with the underlying asset’s price, time to expiration, and volatility.
As the expiration date approaches, the delta of in-the-money options increases towards 1 for
calls, reflecting the increasing likelihood that the option will remain in the money. Conversely,
the delta of out-of-the-money options decreases towards 0 as expiration approaches, reflecting
the decreasing likelihood of the option expiring in the money.

Delta is also used in hedging strategies, such as delta-neutral trading, where the goal is
to offset potential losses in one position with gains in another by maintaining a delta of zero.
This involves adjusting the positions in the underlying asset and options to neutralize the overall
delta of the portfolio.

1.2.2 Gamma (Γ)

Gamma is another key metric from the family of ”Greeks” in options trading, and it measures
the rate of change of an option’s delta in response to a one-dollar change in the price of the
underlying asset. While delta gives an estimate of how the price of an option changes with
movements in the underlying asset, gamma indicates how quickly the delta itself changes. This
measure is crucial for understanding the sensitivity of an option’s price to market movements,
especially for traders managing complex portfolios or employing dynamic hedging strategies [3].

Gamma is particularly significant for options that are near or at the money (ATM), where the
delta is most sensitive to price changes in the underlying asset. A high gamma value indicates
that the delta of the option is highly responsive to changes in the price of the underlying asset.
In practical terms, this means that as the stock price moves, the rate at which the option’s price

2
is expected to change (its delta) will also change rapidly. This can lead to larger than expected
changes in the option’s price, presenting both opportunities and risks for traders.

Gamma is higher for options that are closer to their expiration date. As the time to ex-
piration decreases, the sensitivity of delta to changes in the underlying asset’s price increases,
making accurate prediction of the option’s price movement more critical—and challenging.

Gamma reaches its highest value for ATM options because their delta is most responsive
to price changes in the underlying asset. For options deep in or out of the money, gamma tends
to be lower, indicating that changes in the underlying price have a less dramatic effect on the
option’s delta and, consequently, its price.

For portfolio managers and traders, gamma is crucial for managing the delta of a portfolio. A
portfolio with a high gamma requires more frequent adjustments to maintain a delta-neutral posi-
tion, as small changes in the underlying asset’s price can cause significant changes in delta. This
makes gamma a critical factor in dynamic hedging strategies, where the goal is to neutralize not
just the directional risk (delta) but also the risk related to the rate of change of this risk (gamma).

A high gamma can be both a risk and an opportunity. For traders holding options with
a high gamma, small price movements in the underlying asset can lead to significant profits but
also substantial losses. Therefore, understanding gamma allows traders to better navigate the
markets by anticipating changes in the price behavior of their options.

1.2.3 Theta (Θ)

Theta is another important benchmark belonging to the option Greeks. Theta measures the
sensitivity of an option’s price to the passage of time, indicating how the price of the option
decreases as it approaches its expiration date. Generally speaking, the value of theta is more
negative for options with a shorter period of time until expiration. As an option approaches full
maturity, its theta becomes more pronounced due to the diminishing time value of the option
[4].

Furthermore, the value of theta also depends on the strike price of an option. Theta is
typically highest when the strike price is close to the actual price of the underlying asset, in
other words, within the region of ATM (at-the-money). It then decreases in both directions as
the price of the underlying asset moves either ITM (in-the-money) or OTM (out-of-the-money).
High theta values for ATM options reflect the accelerated time decay as the expiration date
approaches, causing a rapid decrease in the option’s price.

It is important to keep in mind that theta, while a useful measure, represents the theoretical
rate of time decay and assumes that all other factors remain constant. In reality, the actual
time decay can deviate from the projected values due to market fluctuations and changes in
volatility. Therefore, continuous monitoring and recalculation of theta are essential to maintain
accurate and relevant valuations.

2 Theory
The Black-Scholes model provides a theoretical estimate of how an option’s value is affected by
time decay, volatility, and other factors.

3
2.1 Black-Scholes Option Pricing Model
The Black-Scholes formula for the price of a European call option is given by [5] :

C(S, t) = St Φ(d1 ) − Ke−rt Φ(d2 )

where:

– C(S, t) is the price of the call option as a function of the stock price S and time t.

– St is the current stock price.

– Φ(d) is the cumulative distribution function of the standard normal distribution.
– K is the strike price of the option.

– e−rt is the discount factor, where r is the risk-free interest rate and t is the time to
expiration.
– d1 and d2 are calculated as follows:

 
St σ2


ln K + r+ 2 τ
d1 = √
σ τ
√
d2 = d1 − σ τ ,

– σ is the volatility of the stock’s returns.

– τ = T − t.

2.2 Derivation of Black-Scholes Price

Given a European call option with payoff max(ST − K, 0), where ST is the stock price at time T
and K is the strike price. Let C(St , t) be the price of the call option at time t with stock price St .

The stock price dynamics under the Black-Scholes model are given by the stochastic differential
equation (SDE) [6]:

dSt = µSt dt + σSt dWt

where µ is the drift rate, σ is the volatility, and Wt is a Wiener process.

Using Itō’s lemma to find the differential of C(St , t), we have:

1 2 2 ∂2C
 
∂C ∂C ∂C
dC = + µSt + σ St dt + σSt dWt
∂t ∂St 2 ∂St2 ∂St
∂C
To eliminate risk, construct a portfolio Π by shorting ∆ = ∂St shares of stock and holding one
option. The portfolio value is:

Π = C − ∆St
The differential of Π is:

4
 dΠ = dC − ∆dSt
∂C
Substituting dC and dSt , and choosing ∆ = ∂S t
to eliminate the risk (terms involving dWt ):

1 2 2 ∂2C
 
∂C
dΠ = + σ St dt
∂t 2 ∂St2
Since Π is risk-free, it must grow at the risk-free rate r:

dΠ = rΠdt = r(C − ∆St )dt

Equating the expressions for dΠ:

∂C 1 ∂2C ∂C
+ σ 2 St2 2 + rSt − rC = 0
∂t 2 ∂St ∂St
This is the Black-Scholes PDE. For a European call option, solving this PDE with the final
condition C(ST , T ) = max(ST − K, 0) yields the Black-Scholes formula:

C(St , t) = St Φ(d1 ) − Ke−rτ Φ(d2 )

where
 
St σ2


ln K + r+ 2 τ √
d1 = √ , d2 = d1 − σ τ
σ τ
Φ denotes the cumulative distribution function of the standard normal distribution.

In deriving the Black-Scholes formula several assumptions [7][8] have been made, these being
the following:

a. The risk-free interest rate is constant.

b. The price of the asset follows a random walk, in accordance with the findings of Bachelier
[9].
c. No dividends are paid out during the holding period of the option.
d. The option is European, with the only possibility of utilizing the option at its full maturity.
e. Transaction costs do not exist.

f. Variance is constant over the life of the option.

g. Stock prices are log-distributed.
h. There are no arbitrage opportunities.

Yielding from these assumptions is the fact that the future value of the option is solely based
on known values and constants, such as the risk-free rate, the time for holding the option and
the option price at the date of acquirement.

5
2.3 The Greeks for the Black-Scholes
Black-Scholes also derives the Greeks which are critical for managing the risks associated with
options portfolios by quantifying the sensitivity of the option’s price to various underlying
factors.

2.3.1 Delta (∆)

Delta for a European call option can be obtained by differentiating the Black-Scholes formula
with respect to the stock price St [10]. Analytically, it is given by:

∂C
∆= = Φ(d1 )
∂St

2.3.2 Gamma (Γ)

Gamma measures, as mentioned earlier, the rate of change of delta with respect to changes in
the underlying asset price [10]. It can be derived by differentiating delta with respect to St :

∂2C ∂∆ φ(d1 )
Γ= 2 = = √
∂St ∂St St σ τ

where ϕ(d1 ) is the probability density function of the standard normal distribution evaluated at
d1 .

Note: Φ(d) represents the cumulative distribution function of the standard normal distribution,
and φ(d) represents the probability density function of the standard normal distribution.

∂Γ
2.3.3 Derivative of Gamma ( ∂S t
)
In the context of greeks in the Black-Scholes model, the derivative of Γ not widely used. In this
∂Γ
study, further information than what ∆ and Γ gives will be obtained by calculating ∂S t
. Even
though this is not used to hedge the portfolio in particular, this serves one important predictor
once Γ is being predicted.
∂φ(d1 ) ∂φ(d1 )
∂Γ ∂ φ(d1 ) 1 ∂St St − φ(d1 ) ∂St St − φ(d1 )
= √ = √ √ = =
∂St ∂St St σ τ σ τ (St σ τ )2 St2 σ 3 τ 3/2
∂φ 1 K √ ∂φ (d1 ) − φ(d1 )
K
∂St (d1 ) σ τ St St − φ(d1 )
√
 
∂φ ∂φ 1 K σ τ ∂St
= (d1 ) = (d1 ) √ = = .
∂St ∂St σ τ St St2 σ 3 τ 3/2 St2 σ 3 τ 3/2
∂φ
Here, d1 is as defined before and ∂St is the derivative of the function of a standard Gaussian
density.

2.3.4 Theta (Θ)

Given the Black-Scholes formula for a European call option [10]:

C(S, t) = SΦ(d1 ) − Ke−r(T −t) Φ(d2 )

6
where:
 2

S
+ r + σ2 (T − t)


log K
d1 = √ ,
σ T −t
√
d2 = d1 − σ T − t,

and Φ is the cumulative distribution function of the standard normal distribution, σ is the
volatility of the underlying asset.

To find theta, we need the partial derivatives of d1 and d2 with respect to t:


S

 σ2
 
∂d1 ∂  log K + r + 2 (T − t)
= √ 
∂t ∂t σ T −t
 
σ2 3
S
− t) · 21 (T − t)− 2 σ
2


−(r + σ2 ) log K + (r + 2 )(T
= √ +
σ T −t σ(T − t)
 2

S
+ r + σ2 (T − t)
2


−(r + σ2 ) log K
= √ − 3
σ T −t 2σ(T − t) 2
2 √
 
σ2
−(r + σ2 ) T − t − log KS


+ (r + 2 )(T − t)
=
2σ(T − t)
Similarly, the partial derivative of d2 with respect to t is:
∂d2 ∂d1 ∂ √
= − (σ T − t)
∂t ∂t ∂t
2 √
 2

−(r + σ2 ) T − t − log K
S
+ (r + σ2 )(T − t)

1
σ · 21 (T − t)− 2 (−1)
= − √
2σ(T − t) T −t
σ2 √ 
S

σ 2

−(r + 2 ) T − t − log K + (r + 2 )(T − t) σ
= + √
2σ(T − t) 2 T −t
The partial derivative of the Black-Scholes formula with respect to t is:
∂C ∂Φ(d1 ) ∂d1 ∂Φ(d2 ) ∂d2
= −rKe−r(T −t) Φ(d2 ) + S − Ke−r(T −t)
∂t ∂d1 ∂t ∂d2 ∂t
Given Φ′ (d) = ϕ(d), the density function of the normal distribution, we have:
∂C ∂d1 ∂d2
= −rKe−r(T −t) Φ(d2 ) + Sϕ(d1 ) − Ke−r(T −t) ϕ(d2 )
∂t ∂t ∂t
∂C
Θ=−
∂t
Thus, theta quantifies the sensitivity of the option’s price to the passage of time, indicating the
rate at which the price of the option decreases as it approaches its expiration date. This expres-
sion, Θ, can be broken down into three components: the loss due to the approaching expiration
(−rKe−r(T −t) Φ(d2 )), the time decay effect on d1 and d2 (Sϕ(d1 ) ∂d
∂t and −Ke
1 −r(T −t)
ϕ(d2 ) ∂d
∂t ).
2

7
3 Hedging Strategy & Algorithmic Application
The essence of delta-gamma hedging involves adjusting a portfolio in such a way that it is
neutral to both the direction of the market movements (delta-neutral) and the curvature of
how option prices change with those movements (gamma-neutral). This dual neutrality helps
in managing the risks associated with small and incremental price changes in the underlying
asset. Delta-gamma hedging exemplifies the sophistication possible in options trading and
risk management, offering a nuanced approach to stabilizing a portfolio’s value against minor
fluctuations in market prices.

3.1 Delta-Neutrality
The first step in delta-gamma hedging is to neutralize the delta of the portfolio. This is done
by adjusting the holdings of the underlying asset or other derivative instruments so that the
overall delta of the portfolio is as constant as possible. A delta-neutral portfolio is not affected
by small movements in the price of the underlying asset because gains or losses on the options
positions are offset by changes in the value of the underlying holdings.

3.2 Gamma-Neutrality
The second step focuses on neutralizing gamma, ensuring that the delta of the portfolio remains
stable even if the underlying price continues to move. This is crucial because, in a delta-neutral
portfolio, the delta can change with movements in the underlying asset’s price, necessitating
continuous rebalancing. By also achieving gamma neutrality, the portfolio is insulated against
the need for frequent adjustments since the delta will not change significantly with small
movements in the underlying price.

3.3 Delta-gamma hedging

Suppose we are at time t and are locked in with a call option on an underlying asset of price
St . We are unable to sell the option right away and would like to neutralize movements in the
option’s price at future time steps with a portfolio πt = (xt , 1, yt ) where xt , yt are respectively
the number of shares held in the asset and yt the quantity held in another option at time t.
Why is another option needed? Since the gamma of a portfolio is the linear combination of its
gammas and that stocks have constant Γ = 1, we cannot achieve 0 gamma only by shorting
a stock so we would need to hold a certain quantity of another call option for the equation
Γt + yt Γ′t = 0 to hold, where Γ, Γ′ are the gamma’s of respectively our initial option and the one
sold to hedge.

8
 Figure 1: Diagram explaining the time steps in the delta-gamma hedging strategy.

3.4 Machine Learning Algorithms

In this study the purpose is to forecast Γ, which by definition will tell us about ∆. Some of
the algorithms used are non-linear regression (with some type of penalization method), random
forest, gradient boosting algorithms. This will be complemented by some comparisons of the
ARMA-GARCH time series model.

3.4.1 Linear Regression

Linear regression is a statistical method used to model the relationship between a dependent
variable and one or more independent variables by fitting a linear equation to observed data.
The simplest form of linear regression is one with two variables that fits a linear equation of a
slope which intercept the data points. In multiple linear regression, several independent variables
are used to predict the dependent variable.

The linear equation for simple linear regression is represented as:

y = β0 + β1 x + ϵ

Where,
– y is the dependent variable,
– x is the independent variable,
– β0 is the y-intercept,
– β1 is the slope of the line,
– ϵ is the error term.

9
In multiple linear regression, the equation expands to accommodate multiple independent
variables (x1 , x2 , . . . , xn ):

y = β0 + β1 x1 + β2 x2 + . . . + βn xn + ϵ.

The coefficients (β) are estimated during the training process using the least squares method,
aiming to minimize the sum of the squared differences between the observed values and the values
predicted by the linear equation. The algorithms simplicity and comprehensibility are two of its
strength while at the same its simple nature being its shortcoming due an under-performance
for complex data sets with patterns that do not follow linearity.

3.4.2 Random Forest

The random forest machine learning algorithm is an ensemble method with an extension of
bagging. By running multiple regression trees and then averaging over them a forest of regression
trees are created, hence the name of the algorithm. In order to capture different perspectives the
random forest also use random feature selection when creating the trees. Bagging, or Bootstrap
Aggregating, trains an algorithm by creating different subsets of the same original data set in
order to achieve a better performing model. Assume there are B number of data sets that are
independent, yielding B number of trees. Then the aggregated model can be written as,
B
X
fagg (x) = f b (x)
b=1

The primary goal of baggging is to decrease a model’s variance while maintaining low bias. A
potential downfall with bagging when using regression trees is the risk of correlated subsets
which random forest address by inducing randomness. Learning a random forest model occurs
in parallel and large number of B datasets does not lead to overfitting.

3.4.3 Gradient Boosting Machines

Gradient Boosting Machine is an iterative machine learning method with a main idea of
aggregating simple models, referred to as weak learners, together. A weak learner could be
of such nature that it barely outperform random chance such as a decision stump, a shallow
decision tree with only one decision node. By sequentially aggregating the weak learners together
by assigning adaptive weights, it is possible to achieve a more robust and better performing
model. At each iteration the algorithm assess and fit the new model by reweighing focusing on
training data points with high past losses, that is points which the model previously performed
weakly. The general boosting algorithm idea when aggregating B number of simple models
looks as follows:
B
X
fboost (x) = sign αb f b (x)
b=1

where αb = are the different adaptive weights for each simple model f b (x). However, boosting
machines may be prune to overfitting when B, the number of models, increases. To address this
issue of choosing an appropriate number of B models, there exist multiple approaches. One is
to stop when an additional model does not increase the overall model’s performance. Another is
to create a performance threshold, at which the algorithm stops when it is reached. B can thus
be tuned as a hyperparameter. One type of of gradient boosting machines is a method called
Xtreme Gradient Boosting, or XGBoost in short. The aim of a boosting algorithm is to reduce
bias.

10
3.4.4 Support Vector Regression
A Support Vector Regression (SVR) is a variant of a Support Vector Machine (SVM). The aim
of a SVR is similarly to a SVM, to find and create the best fitting hyperplane which maximises
the margin between itself and the data points. However, instead of classification, the data
is regressed and the points that fall within the so called tube of margin are considered to be
accurately predicted. Data points outside of this margin are added to the models error as seen
in algorithms loss function below
(
0 if |y − ŷ| < ϵ
L(y, ŷ) =
|y − ŷ| − ϵ otherwise

Being a regression model it yields a continuous value as an output. The kernel trick can be ap-
plied for SVR just as for SVM to transform the input feature space in order to model and capture
non-linear relationship in the data. The SVR tube does therefore not necessarily have to be linear.

SVRs require hyperparameter tuning one of which is known as the regularization factor c.
A larger parameter c equals to a larger margin which implies that the model is less strict towards
errors, whilst a smaller one decreases the margin making the model more sensitive to errors.
The data points that are on the boundary or outside the ϵ-insensitive tube are the support
vectors and are important in the constituting the model. Relatively speaking, SVR is more
robust towards outliers in comparison to a normal regression, as the primary focus is on data
points close to the margin.

4 Data Extraction, Cleaning and Exploration

This section involves the initial steps in data analysis where raw data is gathered from various
sources, purified to remove inaccuracies or inconsistencies, and then analyzed to discover patterns
or insights. This process is crucial for ensuring the data is accurate and suitable for modeling or
further analysis.

4.1 Data Selection

For numerical preidictors, historical values of our response variable gamma as well as volatility,
stock price, delta, vega, derivative of gamma, volume, volume change (derivative of volume),
time to maturity, inflation rate and risk free interest rate were retrieved daily at the market’s
closing state. Inflation rate as well as risk free interest rate is updated monthly by the extracting
the data from API.

For categorical values, predictors that identified bull or bear market and weather an option was
in, out or at the money were used.

11
 Table 1: Predictor Descriptions and Formulas

Predictors Type Formula Note

Stock Price Numerical Retrieved (Yahoo Fi-
nance)
∂C
Delta Numerical ∂St = Φ(d1 ) ∂C = Change in Call
Option Price, ∂St =
Change in Stock Price
Call Option Price Numerical C(S, t) = St Φ(d1 ) −
Ke−rt Φ(d2 )
2
Gamma Numerical Γ = ∂∂SC2 Variables explained in
t
section 2.3.2
Theta Numerical θ = ∂P
∂t
Volume Numerical Retrieved (Yahoo Fi-
nance)
Volume Change Numerical ∆Volume = Volumen −
Volumen−1
Bull or Bear Categorical Bear < -20% Index de- Over the time horizons
cline, Bull > 20% Index we chose, this variable
increase was most of the time a
column full of 1 so we did
not use it to avoid multi-
collinearity with the con-
stant column.
In Out At Money Categorical In the Money: S−C > 0, S: Stock Price, C: Call
Out of the Money: S − Price. Same remark as
C < 0, At the Money: the one above.
S−C =0
Time to Maturity Numerical Current date − Maturity Calculated
date
Lagged variables Numerical Lagged variables of the
above variables

Below follows a section that explains the reasoning why the chosen predictors where were
explored in the machine learning models, as well as information on their time interval.

12
 Table 2: Predictors and their Relevance and time span

Predictor Relevance Time Interval

Volatility Relevant for determining risk Calculated daily with a periodic-
and pricing of option. ity of 21 days over one year (252
trading days).
Stock price Contains most information about Continuous over time, retrieved
the company, market expecta- daily
tions, trends, performance etc.
Delta Calculates change in option price Continuous over time, retrieved
per change in the underlying as- daily.
set, thus, crucial for delta-gamma
hedging.
Historical gamma Corresponding variable Continuous over time, retrieved
daily.
Volume Volume can indicate the capacity Continuous over time, retrieved
and demand of the market as well daily.
as pricing of options and change
of hedging strategy.
Volume Change Change of volmme can change op- Continuous over time, retrieved
tion pricing hence bring the need daily.
to adjust the hedging.
Bull or Bear Can predict the direction of a Retrieved Daily.
market as a whole.
In/At/Out of Money Can be an indicator of high/low Continuous over time, retrieved
Delta as said in the theory section daily.

4.2 Stock Prices, Options Prices and Greeks

Yahoo Finance offers free access to stock prices over a large time period and which can be
easily loaded in Python. It also has prices for some call options but that cannot be loaded
directly to Python. The Swiss Federal Institute of Technology, EPFL, grants partial access
to WRDS database but it only has samples of options data and contained a significant num-
ber of wrong or missing information. OPRA dataset from Databento is very comprehensive,
maybe a bit too much and it is not entirely free. Polygon.io is also pretty good but it has
lots of missing data. There are also other websites that give live data for options but no
access to previous data. Since options data can be retrieved easily from the theory stated
above, we chose to load Yahoo Finance stock prices and use Black-Scholes model to find
the price of call options with a given maturity, strike at a given moment. Same goes for
Greeks. If the data collected from Yahoo Finance is accurate, it can be confidently concluded
that the options data independently calculated will be free from missing or incorrect information.

The stock prices for Apple from 1st January 2023 to 16th May 2024 and a call option with
strike 185$ and maturity 17th May 2024 may be seen in the following figure:

13
 Figure 2: Price of Apple stock and a call option during a 15-month period.

The Greeks for the same period and same option are as follows. We see that ∆ and Γ both
belong to [0, 1] as they should. The numbers correspond pretty well to the ones we can find in
the website Barchart.

Figure 3: Delta (∆) and Gamma (Γ) for a call option for Apple stock during a 15 month
period.

The historical volatility we chose for stocks comes from Barchart which is calculated over minute
data rather than the daily data Yahoo Finance offers.

We rescale the volume variable such that the regression coefficients are not too high or too small.

For our regression, we want to predict Γ at time t and we start by making up our design
matrix containing the variables we have at t − 1. At the latter time, we know the time to
maturity and all lagged variables such as prices, volume and Greeks. We also include 2 day
lagged variables to look at changes in quantities. Our variables are hence time to maturity,
Γt−1 , ∆t−1 , Ct−1 , St−1 , Θt−1 , Vt−1 , Γt−2 , ∆t−2 , Ct−2 , St−2 , Θt−2 , Vt−2 . We drop variables that
will have NaN values for lagged variables.

14
We move on to a closer look to our variables. Our dependent variable Γ has a skewness
of 4.93 and it is really right-skewed as shown by the histogram below so we decide to apply a
log-transformation to it, which makes it clearly better.

Figure 4: Histogram of transformed Gamma to unskew data.

We also look at histograms and boxplots of the independent variables. Some of them are also
right-skewed and need log-transformation. The volume is such an example as shown below.

Figure 5: Histogram and boxplot of Lagged Volume before and after log-transformation

We then move on to some bivariate analysis. We plot variables against each other and have
some interesting results. We see that some clear relations appear with Γ. The whole picture
being too big to export, we only give examples of interesting plots. Here is Gamma against time
to maturity, we clearly observe some -log shape.

15
 Figure 6: Plot of Log Gamma against time to maturity

Some variables don’t seem to have predictive power which matches our expectations about them.
It is the case of volume as we can see in the plot below.

Figure 7: Plot of Log Gamma against lagged volume

There are variables that do not change a lot between time t − 1 and t − 2 so to avoid a case of
near-multicollinearity, where det(X T X) ≈ 0, we should include one or another, not both.

16
 Figure 8: Plot of Log 1 Day lagged gamma vs Log 2 Days lagged gamma

Finally, we have a look at a correlation heatmap of our variables. We see that Γ at time t is very
highly correlated with its value at time t − 1. We also notice as said before that the volume is
poorly correlated with Γ so we decide not to keep it for the regression. We can also see that ∆
has zero correlation with Γ, we could remove it as we did for volume but it is intuitively linked
with Γ in options theory and encodes information about underlying’s price so we keep it for now.

Figure 9: Plot of Log 1 Day lagged gamma vs Log 2 Days lagged gamma

17
5 Implementation
This section includes process of designing, training, and testing machine learning algorithms
based on the cleaned and explored data to make predictions or automate decision-making. This
stage involves selecting appropriate models, tuning parameters, and validating the model’s
performance to ensure it meets the specified objectives effectively.

5.1 Regression
We now delve into the implementation of a linear regression model to predict Gamma at a
given time. We must be careful though, as Delta is bounded between 0 and 1, the variation of
delta, i.e., Gamma, is also inherently bounded. However, traditional linear regression techniques
may overlook this constraint, potentially leading to predictions that exceed the feasible range.
To address this challenge, if it arises, we will enforce the bounds of Gamma, by truncating
predictions beyond the feasible range or by mapping them to the nearest feasible boundary (0
or 1).

We choose to implement a backward selection to find the right model. We start with a
model containing all the regressors and removing one at a time to maximize or minimize the
following metrics : R2 , AIC and BIC. The latter procedure leads us to the following table where
we removed the worst-performing variable, i.e. the one where the metrics were behaving best.

Table 5.1.1 Metrics of the models through backward selection

Regressors removed R2 AIC BIC
None 0.8256 -1492.579 -1462.764
Lagged Option Price 0.8294 -1494.562 -1468.474
2 Days Lagged Stock Price 0.8278 -1494.935 -1472.574
Lagged Delta 0.8472 -1491.415 -1472.782
Lagged Stock Price 0.8467 -1493.307 -1478.400

Table 5.1.2 Linear regression results of last model

Dependent variable : log(Γt ) (1)

log(Γt−1 ) 0.93370 ∗∗∗
(0.019)

log(T ) -0.04378 ∗∗∗

(0.011)

Θt−1 0.002 ∗
(0.001)

Constant -0.29641 ∗∗∗

(0.093)
Out-of-sample R2 0.8467
Number of observations (in-sample) 307
Number of observations (out-of-sample) 35
Standard errors in parentheses
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01

18
We now plot residuals against the variables we kept to see if there is some missing form that
was not captured.

Figure 10: Plot of Log Gamma against lagged volume

We see no significant pattern which is a good sign. We also check for heteroskedasticity by
plotting residuals against fitted values.

Figure 11: Plot of residuals against fitted values

The plot looks fine. If we compare our predictions with real data, it is actually pretty close.

19
5.2 Delta-Gamma Predictions
Suppose we are at date i − 1 and want a portfolio where we have a hedged position at date i.
We have a prediction for the next Gamma values given by :
ˆ 2
Γ̂i = eln Γi eσ̂ /2

We still need predictions for the next Delta values. We use a linear approximation the following
way:
∆i ≈ ∆i−1 + Γi−1 · (Si − Si−1 )
Γi − Γi−1
3 ≈ (Si − Si−1 )
( ∂∂SC )i−1
We then solve the linear system in figure 1 to get the positions we need. Here is an example.
Suppose its 16th April 2024 and you are given call options from Apple stock with strike 185
and maturity 04th June 2024. We would like to hedge them with call options of Microsoft stock
with strike 410. Volatility details and examples of strike prices can be found either on a excel
given of on Yahoo Finance. Suppose you want to hedge the options until the 23rd May. Here
are the positions you need to take on the market to be delta-gamma hedged:

Figure 12: Daily positions on the market to be delta-gamma hedged.

20
6 Alternative Methods Results
This is a section dedicated to presenting the outcomes of models that did not perform as
expected, providing insights into their limitations and areas for improvement.

6.1 Random Forest

Table 3: Feature Importances

Feature Importance
log(Γt−1 ) 0.468894
log(T ) 0.402111
Θt−1 0.114102
Pricet−1 0.005920
Ct−2 0.004195
∆t−1 0.002627
Ct−1 0.002151
Mean Squared Error: 0.323615
R2 Score: -0.07625

Figure 13: Feature importances of Random forest model

21
6.2 SVR

Table 4: SVR Model Evaluation with best parameters

Parameter Value
C 1
Gamma 0.01
Epsilon 0.1
Kernel linear
Mean Squared Error 0.07667
R2 Score 0.74502

6.3 XGBoost

Table 5: XGBoost Model Evaluation with Best Parameters

Parameter Value
Regressor XGBoost
Number of Estimators 200
Learning Rate 0.1
Max Depth 3
Min child weight 1
Subsample 1
Colsample Bytree 0.7
Mean Squared Error 0.32122
R2 Score -0.06830

7 Methods comparison
We will now compare the different machine learning models used and determine which one
is the most performant based on two metrics:R2 score and the mean squared error. A lower
MSE indicates better performance, while a higher R2 score (closer to 1) signifies a better fit
to the data. It corresponds to the proportion of the variance in the dependent variable that is
predictable from the independent variables.

First, here is a summary table of the results:

Linear Regression Random forest SVR XGBoost

Mean Squared Error 0.04612 0.323615 0.07667 0.32122
Out of sample R2 0.8467 -0.07625 0.74502 -0.06830

Table 6: Summary table of the results

Let’s first analyze the results of each model:

• Linear regression: as we can see in Table 6, the mean squared error is relatively low
and has pretty good R2 , indicating good performance while having only 4 variables. For
prediction purposes, the model does a good job, the best of the 4 methods, but we should
watch out for heteroskedasticity depending on stocks.

22
 • Random Forest: The random forest model has a significantly higher mean squared error
and a negative R2 value, indicating poor performance. The negative R2 suggests that the
model is worse than a horizontal line predicting the mean of the target variable. This could
be due to overfitting on the training data or insufficient tuning of the hyperparameters.
The complexity of random forest, with its numerous decision trees, might not be suitable
for this dataset without proper parameter tuning.
• Support Vector Regression (SVR): This method has an MSE that is slightly higher than
linear regression but still relatively low compared to the other models. The R2 value is
also fairly high, suggesting that SVR performs well on this dataset. The linear kernel used
in SVR gives it similar results to linear regression, though it doesn’t outperform linear
regression in this case.
• XGBoost: Similar to the random forest, XGBoost has a high mean squared error and
a negative R2 value, indicating poor predictive performance. XGBoost is a powerful
algorithm, but its performance heavily relies on the correct tuning of hyperparameters. In
this instance, it seems to suffer from either overfitting or inappropriate hyperparameters.

8 Conclusion
In summary, the linear regression model outperforms the other models in this scenario, both in
terms of mean squared error and R2 . Random Forest and XGBoost show poor performance,
potentially due to inadequate parameter tuning or overfitting. Support Vector Regression, while
performing well, does not surpass linear regression. The results suggests that using these linear
regressing algorithms would be reasonably feasible for predicting gamma and thus create a
Delta-Gamma neutral portfolio in the market.

For future improvement, having access to a better database would be beneficial since the
data available from Yahoo Finance is limited. Furthermore, due to the relative underperfor-
mance of random forest and XGBoost, a more deliberate parameter tuning for these alternative
methods would be well in place. Finally, an implementation of a more robust linear model to
potentially remove Theta from the regressors would potentially be desirable since heteroskedas-
ticity may arise (depending on stocks), resulting in a falsely low p-value.

For further research or future projects, a next step would be to start a real Delta-Gamma
hedging strategy in the market.

23
References
[1] James Chen. (2024) What are Options? Types, Spreads, Example, and Risk Metrics.
Investopedia. https://www.investopedia.com/terms/o/option.asp

[2] Jhon Summa.(2024) Options Trading Strategies: Understanding Position Delta. Investo-
pedia. https://www.investopedia.com/articles/optioninvestor/03/021403.asp
[3] James Chen.(2024) What Is Gamma in Investing and How Is It Used?. Investopedia.
https://www.investopedia.com/terms/g/gamma.asp
[4] James Chen.(2024) Theta: What It Means in Options Trading, With Examples. Investo-
pedia. https://www.investopedia.com/terms/t/theta.asp
[5] Adam Hayes.(2024) Black-Scholes Model: What It Is, How It Works, and Options
Formula. Investopedia. https://www.investopedia.com/terms/b/blackscholes.asp
[6] Scott Guernsey.(2013).AN INTRODUCTION TO THE BLACK-SCHOLES PDE
MODEL.University of new Mexico. https://math.unm.edu/~nitsche/mctp/reus/
proposals/2013prop_Guernsey.pdf
[7] Black, F., & Scholes, M. (1973). The Pricing of Options and Corporate Liabilities. The
Journal of Political Economy, 81(3), 637–654. https://doi.org/10.1086/260062
[8] Merton, R. C. (1973). Theory of Rational Option Pricing. The Bell Journal of Economics
and Management Science, 4(1), 141–183. https://doi.org/10.2307/3003143
[9] Bachelier, L. (1900). Théorie de la spéculation. Annales Scientifiques de l’École Normale
Supérieure, 17, 21–86. https://doi.org/10.24033/asens.476
[10] Martin Haugh.(2016).The Black-Scholes Model.Foundations of Financial Engineer-
ing, University of Columbia. https://www.columbia.edu/~mh2078/FoundationsFE/
BlackScholes.pdf
[11] In Appendix: Prashant Sharma. (2024) Stock Market Prediction Us-
ing Machine Learning https://www.analyticsvidhya.com/blog/2021/10/
machine-learning-for-stock-market-prediction-with-step-by-step-implementation/

24
A Appendix: Alternative approach
One can say : ”Well, since the formula for Gamma is known, what do we actually need to
calculate tomorrow’s Gamma?”

Γ satisfies the following equation:

−d21 √√
log(Γ) = − log(Sσ t 2π)
2
2 2
− log(S/K)2 log(S/K)(r + σ2 ) (r + σ2 )2 t √ √
= 2
− 2
+ 2
− log(S t) − log(σ 2π)
2σ t σ σ
Then the only thing we do not know to calculate tomorrow’s Γ is the stock price S. If we have a
good prediction for the stock price, then we have a good prediction for Γ. We try to predict the
stock using an arbitrary LSTM model (its parameters are in the code). Here the input is not
the variables used in the methods above but the classic Yahoo Finance dataset. Why? Because
the However, as we can see below, sometimes the predicted values can be far from the real ones
implying big errors in the Gamma prediction which is usually a small quantity.

Figure 14: Apple stock price prediction by LSTM