Advanced Exercises in Physics and Quantitative Finance

Introduction
This document presents a collection of 30 intermediate to advanced exercises
designed to test understanding and application of concepts spanning classical
physics, statistical mechanics, dynamical systems, stochastic calculus, time series
analysis, and econophysics. The exercises draw upon quantitative models and
analytical techniques often employed in scientific research and financial modeling.
Approximately half of the exercises focus on foundational scientific principles, while
the other half explore their application, or the application of related mathematical
tools, to the domain of stock price analysis and prediction.

Each exercise includes a detailed worked solution, providing step-by-step derivations,

calculations, and interpretations. The problems are designed to challenge the user to
move beyond rote memorization and engage with the underlying assumptions,
limitations, and implications of the models and methods presented. This collection
serves as a resource for students, researchers, and practitioners seeking to deepen
their expertise in these interconnected fields.

Index of Exercises
The following table provides a structured overview of the exercises contained within
this document, allowing for targeted selection based on topic, domain, skills tested,
and difficulty level.

Exercise # Primary Relevant Domain Key Skills Difficulty

Topic/Conc Source Tested
ept Material

1 1D Heat S_P1 Physics/Scie PDE Solving Intermediate

Equation nce (Separation
Solution of Variables)
(Dirichlet
BC)

2 1D Heat S_P1 Physics/Scie PDE Solving Intermediate

Equation nce (Separation
Solution of Variables)
(Neumann
BC)
3 Impact of S_P1 Physics/Scie Model Advanced
Non-Consta nce Interpretatio
nt Thermal n, PDE
Diffusivity Analysis

4 Ising Model: S_P3 Physics/Scie Statistical Intermediate

Mean nce Mechanics,
Magnetizatio Approximati
n (High on Methods
Temp)

5 Ising Model: S_P3 Physics/Scie Statistical Intermediate

Mean nce Mechanics
Magnetizatio
n (Low
Temp)

6 Logistic S_P2 Physics/Scie Dynamical Intermediate

Map: Fixed nce Systems
Points and Analysis
Stability

7 Logistic S_P2 Physics/Scie Computation Intermediate

Map: nce al
Numerical Simulation,
Simulation & Dynamical
Bifurcation Systems

8 Lyapunov S_P2, S_C1 Physics/Scie Numerical Advanced

Exponent nce Methods,
Estimation Chaos
(Logistic Theory
Map)

9 Interpretatio S_P2, S_C1 Physics/Scie Conceptual Intermediate

n of Positive nce Understandi
Lyapunov ng (Chaos),
Exponent Model
Interpretatio
n

10 Fourier S_P4 Physics/Scie Signal Intermediate

Transform nce Processing,
Application Fourier
 to Physical Analysis
Signal

11 Interpreting S_P4 Physics/Scie Signal Intermediate

Frequency nce Interpretatio
Spectrum of n, Physical
Physical Modeling
Signal

12 Simplified (Implied by Physics/Scie Fluid Advanced

Navier-Stoke Outline) nce Dynamics,
s: Couette PDE Solving
Flow

13 Network (Implied by Physics/Scie Network Intermediate

Theory: Outline) nce Analysis,
Degree Statistical
Distribution Analysis

14 Entropy (Implied by Physics/Scie Statistical Intermediate

Calculation Outline) nce Mechanics,
(Simple Thermodyna
System) mics

15 Boltzmann (Implied by Physics/Scie Statistical Intermediate

Distribution Outline) nce Mechanics
Application

16 GBM: S_S5 Stock Stochastic Intermediate

Expected Prediction Calculus,
Value of Model
Stock Price Properties

17 GBM: S_S5 Stock Stochastic Intermediate

Variance of Prediction Calculus,
Stock Price Model
Properties

18 GBM: S_S5 Stock Stochastic Intermediate

Expected Prediction Calculus,
Value & Financial
Variance of Metrics
Log-Returns
19 GBM Path S_S5, S_C2 Stock Monte Carlo Intermediate
Simulation Prediction Simulation,
(Euler-Maruy Stochastic
ama) Processes

20 Ito's Lemma S_S7, S_S5 Stock Stochastic Advanced

Application Prediction Calculus
(log(S_t) for (Ito's
GBM) Lemma)

21 Discussion: S_S5 Stock Model Intermediate

GBM Prediction Critique,
Limitations Financial
(Fat Tails) Data
Analysis

22 GARCH(1,1) S_S6 Stock Time Series Intermediate

Volatility Prediction Analysis,
Forecasting Volatility
Modeling

23 GARCH(1,1) S_S6 Stock Time Series Intermediate

Uncondition Prediction Analysis,
al Variance Model
Properties

24 Interpretatio S_S6 Stock Model Intermediate

n of GARCH Prediction Interpretatio
Parameters n, Volatility
(Persistence) Dynamics

25 Stationarity S_S8 Stock Time Series Intermediate

Testing Prediction Analysis,
Concept Hypothesis
(e.g., ADF) Testing

26 Black-Schole S_S1 Stock PDE Advanced

s PDE Prediction Analysis,
Derivation Financial
Step (Heat Derivatives
Eq. Analogy)

27 Critique of S_S1, S_S5 Stock Model Advanced

 Black-Schole Prediction Critique,
s Heat Cross-Discip
Equation linary
Analogy Analysis

28 ABM S_S3, S_S4 Stock Agent-Based Intermediate

Simulation Prediction Modeling,
Analysis: Simulation
Strategy Analysis
Switching
Frequency

29 ABM S_S3, S_S4 Stock Agent-Based Advanced

Simulation Prediction Modeling,
Analysis: Pattern
Identifying Recognition
Market
Stress

30 Machine S_S9 Stock Machine Advanced

Learning Prediction Learning
Model Concepts,
Formulation Model
(SVM for Specification
Prediction)

Part 1: Exercises in Physics and Scientific Phenomena

This section includes exercises based on fundamental physical laws, statistical
mechanics concepts, and dynamical systems theory, often drawing parallels or using
methodologies later applied in financial contexts.

Subsection 1.1: Thermodynamics and Statistical Mechanics

Exercise 1: 1D Heat Equation Solution (Dirichlet BC)
●​ Problem: Solve the one-dimensional heat equation ∂t∂u​=α∂x2∂2u​for a rod of
length L, where u(x,t) is the temperature at position x and time t, and α is the
constant thermal diffusivity. The boundary conditions are fixed at zero
temperature: u(0,t)=0 and u(L,t)=0 for t>0. The initial temperature distribution is
given by u(x,0)=f(x)=sin(L3πx​)+0.5sin(L5πx​).
●​ Worked Solution:
1.​ Assume Separation of Variables: Let u(x,t)=X(x)T(t). Substituting into the
 PDE gives X(x)T′(t)=αX′′(x)T(t).
2.​ Separate: Divide by αX(x)T(t) to get αT(t)T′(t)​=X(x)X′′(x)​. Since the left side
depends only on t and the right side only on x, both must equal a constant,
say −λ.
3.​ Solve Spatial Equation: X′′(x)+λX(x)=0. With boundary conditions X(0)=0
and X(L)=0, the non-trivial solutions occur for λn​=(Lnπ​)2 where n=1,2,3,...,
and the corresponding eigenfunctions are Xn​(x)=sin(Lnπx​).
4.​ Solve Temporal Equation: T′(t)+αλn​T(t)=0. The solution is
Tn​(t)=Cn​e−αλn​t=Cn​e−α(nπ/L)2t.
5.​ General Solution: The general solution is a superposition:
u(x,t)=∑n=1∞​Cn​e−α(nπ/L)2tsin(Lnπx​).
6.​ Apply Initial Condition: At t=0, u(x,0)=f(x)=∑n=1∞​Cn​sin(Lnπx​). This is a
Fourier sine series expansion of f(x). By comparing the given
f(x)=sin(L3πx​)+0.5sin(L5πx​) with the series, we can identify the coefficients
directly.
7.​ Identify Coefficients: We see that C3​=1, C5​=0.5, and all other Cn​=0.
8.​ Final Solution: Substitute the coefficients back into the general solution:
u(x,t)=1⋅e−α(3π/L)2tsin(L3πx​)+0.5⋅e−α(5π/L)2tsin(L5πx​)

Exercise 2: 1D Heat Equation Solution (Neumann BC)

●​ Problem: Solve the one-dimensional heat equation ∂t∂u​=α∂x2∂2u​for a rod of
length L with insulated ends. The boundary conditions are ∂x∂u​(0,t)=0 and
∂x∂u​(L,t)=0 for t>0. The initial temperature distribution is u(x,0)=f(x)=cos(L2πx​).
●​ Worked Solution:
1.​ Separation of Variables: As in Exercise 1, we arrive at X′′(x)+λX(x)=0 and
T′(t)+αλT(t)=0.
2.​ Solve Spatial Equation (Neumann BC): The boundary conditions are now
X′(0)=0 and X′(L)=0.
■​ If λ=0, X′′(x)=0⟹X(x)=Ax+B. X′(x)=A. X′(0)=0⟹A=0. X′(L)=0 is also
satisfied. So X0​(x)=B (a constant) is a solution for λ0​=0.
■​ If λ>0, let λ=k2. X(x)=Acos(kx)+Bsin(kx). X′(x)=−Aksin(kx)+Bkcos(kx).
■​ X′(0)=Bk=0. Since k=0, we must have B=0.
■​ X′(L)=−Aksin(kL)=0. For non-trivial solutions (A=0), we need sin(kL)=0.
This means kL=nπ for n=1,2,3,....
■​ So, kn​=Lnπ​, and λn​=kn2​=(Lnπ​)2. The corresponding eigenfunctions are
Xn​(x)=cos(Lnπx​) for n=1,2,3,....
3.​ Solve Temporal Equation: For λ0​=0, T′(t)=0⟹T0​(t)=C0​(constant). For
λn​=(nπ/L)2, Tn​(t)=Cn​e−α(nπ/L)2t.
4.​ General Solution: The general solution is a superposition including the n=0
 term: u(x,t)=C0​+n=1∑∞​Cn​e−α(nπ/L)2tcos(Lnπx​)
5.​ Apply Initial Condition: At t=0, u(x,0)=f(x)=C0​+∑n=1∞​Cn​cos(Lnπx​). This is a
Fourier cosine series.
6.​ Identify Coefficients: Comparing with the given f(x)=cos(L2πx​), we see that
C2​=1 and all other coefficients (C0​and Cn​for n=2) are zero.
7.​ Final Solution: u(x,t)=e−α(2π/L)2tcos(L2πx​)
○​ Note: The average temperature remains constant over time with insulated
boundaries, consistent with C0​=0 here because the initial average
temperature was zero. If f(x) had a non-zero average, C0​would be that
average value.

Exercise 3: Impact of Non-Constant Thermal Diffusivity

●​ Problem: Consider the heat equation again, but now assume the thermal
diffusivity α is not constant, but depends on temperature, α=α(u). Write down the
modified 1D heat equation. Briefly discuss why solving this non-linear equation is
significantly more challenging than the constant α case solved in Exercises 1 and
2.
●​ Worked Solution:
1.​ Modified Heat Equation: The heat flux is often modeled as q=−k∂x∂u​, where
k is the thermal conductivity. The conservation of energy leads to
∂t∂u​=−cρ1​∂x∂q​, where c is specific heat and ρ is density. Substituting q gives
∂t∂u​=cρ1​∂x∂​(k∂x∂u​). The thermal diffusivity is defined as α=k/(cρ).
2.​ If k (and thus α) depends on u, we cannot pull it out of the derivative. The
equation becomes: ∂t∂u​=∂x∂​(α(u)∂x∂u​) Alternatively, expanding the
derivative: ∂t∂u​=α(u)∂x2∂2u​+dudα​(∂x∂u​)2
3.​ Challenges:
■​ Non-linearity: The equation is now non-linear because the coefficient
α(u) depends on the solution u itself, and the expanded form contains
terms like (∂x∂u​)2.
■​ Superposition Principle Fails: The principle of superposition, which was
crucial for constructing the general solution from basis functions
(sines/cosines) in Exercises 1 and 2, does not hold for non-linear
equations. We cannot simply add solutions together to get another
solution.
■​ Analytical Solutions Rare: Standard techniques like separation of
variables typically fail for such non-linear PDEs. Analytical solutions are
rare and usually only exist for very specific forms of α(u) or specific
boundary/initial conditions, often requiring advanced transformations.
■​ Numerical Methods: Solving generally requires numerical methods (e.g.,
 finite difference, finite element methods), which introduce their own
complexities like stability and convergence analysis. The dependence of α
on u needs to be handled carefully within the numerical scheme.
■​ The reliance on idealized models with constant coefficients simplifies
analysis considerably, but introducing realistic dependencies, such as
temperature-dependent diffusivity, reveals the mathematical complexities
inherent in more accurate physical descriptions.

Exercise 4: Ising Model: Mean Magnetization (High Temp)

●​ Problem: Consider a simplified 1D Ising model with N spins si​=±1, zero external
magnetic field (H=0), and nearest-neighbor interaction energy E=−J∑i=1N−1​si​si+1​.
In the high-temperature limit (T→∞, or β=1/(kB​T)→0), estimate the average
magnetization per spin, m=⟨M⟩/N, where M=∑i​si​.
●​ Worked Solution:
1.​ Partition Function: The partition function is Z=∑{s}​e−βE, where the sum is
over all 2N spin configurations.
2.​ Average Magnetization: The average total magnetization is
⟨M⟩=Z1​∑{s}​(∑i​si​)e−βE.
3.​ High-Temperature Limit: As T→∞, β→0. Therefore, e−βE→e0=1 for any
finite energy E.
4.​ Simplified Partition Function: In this limit, Z≈∑{s}​1=2N (the total number of
states).
5.​ Simplified Average Magnetization: ⟨M⟩≈2N1​{s}∑​(i∑​si​)
6.​ Symmetry: Consider the sum ∑{s}​(∑i​si​). For every configuration {s1​,...,sN​} with
total magnetization M, there is a corresponding configuration {−s1​,...,−sN​} with
total magnetization −M. Since the energy term is negligible (e−βE≈1), these
configurations contribute equally in magnitude but oppositely in sign to the
sum.
7.​ Alternatively, consider the average of a single spin ⟨sj​⟩: ⟨sj​⟩≈2N1​{s}∑​sj​For any
fixed j, half the configurations have sj​=+1 and half have sj​=−1. The sum
∑{s}​sj​=0.
8.​ Result: Therefore, ⟨M⟩=∑i​⟨si​⟩≈∑i​0=0.
9.​ Magnetization per Spin: m=⟨M⟩/N≈0/N=0.
○​ Interpretation: At very high temperatures, thermal energy overwhelms the
interaction energy J. Each spin behaves almost independently, randomly
pointing up or down with equal probability. There is no long-range order, and
the net magnetization averages to zero. This aligns with the concept that
ordered phases (like ferromagnetism) typically exist only below a critical
 temperature.

Exercise 5: Ising Model: Mean Magnetization (Low Temp)

●​ Problem: Consider the same 1D Ising model as in Exercise 4, but now in the
low-temperature limit (T→0, or β→∞). Assume the interaction coupling J>0
(ferromagnetic). Estimate the average magnetization per spin, m.
●​ Worked Solution:
1.​ Energy: E=−J∑i=1N−1​si​si+1​. Since J>0, the energy is minimized when adjacent
spins are aligned (si​si+1​=1).
2.​ Ground States: The lowest energy states (ground states) are the
configurations where all spins are aligned:
■​ All spins up: si​=+1 for all i. E=−J(N−1). M=+N.
■​ All spins down: si​=−1 for all i. E=−J(N−1). M=−N.
3.​ Low-Temperature Limit: As T→0, β→∞. The Boltzmann factor e−βE
becomes highly dominant for the lowest energy states. Any state with higher
energy E′>Eground​will have a contribution e−βE′ that is vanishingly small
compared to e−βEground​.
4.​ Partition Function: In this limit, the partition function is dominated by the
ground states: Z={s}∑​e−βE≈e−βEground​+e−βEground​=2eβJ(N−1) (Assuming
only the two ground states contribute significantly).
5.​ Average Magnetization: The average magnetization sum is also dominated
by the ground states:
{s}∑​Me−βE≈(+N)e−βEground​+(−N)e−βEground​=NeβJ(N−1)−NeβJ(N−1)=0
6.​ Wait - Spontaneous Magnetization: The above calculation suggests ⟨M⟩=0.
However, this is due to summing over both ground states equally. In a real
physical system below the critical temperature, spontaneous symmetry
breaking occurs. The system will settle into one of the ground states.
7.​ Alternative View (Spontaneous Symmetry Breaking): At T=0, the system
must be in a ground state to minimize energy. If it settles into the all-spins-up
state, M=+N. If it settles into the all-spins-down state, M=−N. The magnitude
of the magnetization is ∣M∣=N.
8.​ Magnetization per Spin: The magnitude of the magnetization per spin is
∣m∣=∣M∣/N=N/N=1.
○​ Interpretation: At very low temperatures, the system minimizes its energy by
aligning all spins. This results in a spontaneous magnetization, with ∣m∣=1.
The sign (+1 or -1) depends on initial conditions or infinitesimal external fields
during cooling. Although the strict calculation averaging over both ground
states gives zero, the physically relevant quantity in the context of phase
transitions is the spontaneous magnetization magnitude, which is 1 at T=0.
 (Note: For a true 1D Ising model, the critical temperature is actually Tc​=0, so
spontaneous magnetization only occurs at T=0. Higher-dimensional models
have Tc​>0).

Subsection 1.2: Dynamical Systems and Chaos

Exercise 6: Logistic Map: Fixed Points and Stability
●​ Problem: Consider the logistic map xn+1​=rxn​(1−xn​), a simple model exhibiting
chaotic behavior. Find the fixed points of the map and determine their stability as
a function of the parameter r (assume 0≤r≤4 and 0≤x≤1).
●​ Worked Solution:
1.​ Find Fixed Points: Fixed points x∗ satisfy x∗=f(x∗)=rx∗(1−x∗).
■​ One solution is x∗=0.
■​ If x∗=0, we can divide by x∗: 1=r(1−x∗).
■​ 1/r=1−x∗⟹x∗=1−1/r.
2.​ Fixed Points Summary: The fixed points are x1∗​=0 and x2∗​=1−1/r. Note that
x2∗​is physically relevant (i.e., 0≤x2∗​≤1) only when r≥1.
3.​ Stability Analysis: Stability is determined by the derivative of the map
function, f′(x)=dxd​[rx(1−x)]=r−2rx. A fixed point x∗ is stable if ∣f′(x∗)∣<1.
4.​ Stability of x1∗​=0:
■​ f′(0)=r−2r(0)=r.
■​ The fixed point x1∗​=0 is stable if ∣r∣<1. Since r≥0, it's stable for 0≤r<1. It
loses stability at r=1.
5.​ Stability of x2∗​=1−1/r: (This point exists for r≥1)
■​ f′(x2∗​)=r−2r(1−1/r)=r−2r+2=2−r.
■​ The fixed point x2∗​is stable if ∣2−r∣<1.
■​ This inequality means −1<2−r<1.
■​ 2−r<1⟹r>1.
■​ 2−r>−1⟹r<3.
■​ So, x2∗​=1−1/r is stable for 1<r<3.
6.​ Summary of Stability:
■​ For 0≤r<1: x1∗​=0 is stable. x2∗​is not positive.
■​ For r=1: x1∗​=0 and x2∗​=0 coincide and are marginally stable.
■​ For 1<r<3: x1∗​=0 is unstable. x2∗​=1−1/r is stable.
■​ For r=3: x1∗​=0 is unstable. x2∗​=1−1/3=2/3 is marginally stable
(f′(2/3)=2−3=−1). This is where the first period-doubling bifurcation
occurs.
■​ For r>3: Both fixed points are unstable. The system exhibits
period-doubling bifurcations leading to chaos.
Exercise 7: Logistic Map: Numerical Simulation & Bifurcation
●​ Problem: Numerically simulate the logistic map xn+1​=rxn​(1−xn​) for r=2.8, r=3.2,
and r=3.9. Start with x0​=0.5. Iterate 100 times, discard the first 50 iterations
(transients), and plot the values of xn​for the remaining 50 iterations for each r.
Describe the qualitative behavior observed in each case.
●​ Worked Solution:
○​ Method: Use a simple loop in a programming language (like Python) to
compute the iterations. Store the last 50 values for plotting.
○​ Code Snippet (Conceptual Python):
Python​
import numpy as np​
import matplotlib.pyplot as plt​
​
def logistic_map(r, x0, n_iter, n_skip):​
x = np.zeros(n_iter)​
x = x0​
for i in range(n_iter - 1):​
x[i+1] = r * x[i] * (1 - x[i])​
return x[n_skip:]​
​
r_values = [2.8, 3.2, 3.9]​
x0 = 0.5​
n_iter = 100​
n_skip = 50​
​
plt.figure(figsize=(12, 4))​
for i, r in enumerate(r_values):​
final_x = logistic_map(r, x0, n_iter, n_skip)​
plt.subplot(1, len(r_values), i + 1)​
plt.plot(range(n_skip, n_iter), final_x, 'k.', markersize=2) # Plot points​
# Or plot lines plt.plot(range(n_skip, n_iter), final_x, 'b-')​
plt.title(f'r = {r}')​
plt.xlabel('Iteration (n)')​
plt.ylabel('x_n')​
plt.ylim(0, 1)​
plt.tight_layout()​
plt.show()​

○​ Expected Observations & Qualitative Description:

■​ r = 2.8: (Region 1<r<3) The plot should show that after the initial transient,
xn​converges to a single fixed value. From Exercise 6, this stable fixed
point is x∗=1−1/2.8≈0.643. The plot will show points clustered tightly
 around this value. Behavior: Convergence to a stable fixed point.
■​ r = 3.2: (Region 3<r<1+6​≈3.449) The plot should show that xn​does not
converge to a single value but oscillates persistently between two distinct
values. This is a stable period-2 cycle, occurring after the first
period-doubling bifurcation at r=3. Behavior: Stable period-2 cycle.
■​ r = 3.9: (Region r>3.57, deep into the chaotic regime) The plot should show
points scattered apparently randomly across a range of values within (0,1),
with no discernible simple pattern or convergence. The values will appear
erratic and sensitive to the exact iteration number. Behavior: Chaotic
dynamics.
○​ Interpretation: This simulation demonstrates how changing a single
parameter (r) in a simple non-linear dynamical system can lead to qualitatively
different behaviors, including stable points, periodic cycles, and chaos,
illustrating the concept of bifurcations.

Exercise 8: Lyapunov Exponent Estimation (Logistic Map)

●​ Problem: Estimate the Lyapunov exponent λ for the logistic map
xn+1​=f(xn​)=rxn​(1−xn​) at r=3.9, using the definition
λ=limN→∞​N1​∑n=0N−1​ln∣f′(xn​)∣. Use N=1000 iterations, starting from x0​=0.5,
and discard the first 100 iterations for the calculation to allow transients to decay.
●​ Worked Solution:
1.​ Map Function: f(x)=rx(1−x) with r=3.9.
2.​ Derivative: f′(x)=r−2rx=3.9−7.8x.
3.​ Iteration: Generate a sequence x0​,x1​,...,xN​using the map xn+1​=f(xn​). Start
with x0​=0.5. Let Ntotal​=1100.
4.​ Lyapunov Summation: Calculate the sum required, discarding the first
Nskip​=100 terms: S=n=Nskip​∑Ntotal​−1​ln∣f′(xn​)∣=n=100∑1099​ln∣3.9−7.8xn​∣
The number of terms in the sum is N=Ntotal​−Nskip​=1000.
5.​ Calculate Exponent: The estimated Lyapunov exponent is
λ≈N1​S=10001​∑n=1001099​ln∣3.9−7.8xn​∣.
6.​ Numerical Implementation (Conceptual Python):​
Python​
import numpy as np​
​
r = 3.9​
x0 = 0.5​
n_total = 1100​
n_skip = 100​
​
 x = x0​
log_f_prime_sum = 0.0​
# Iterate and calculate sum​
for n in range(n_total):​
if n >= n_skip:​
f_prime = r - 2 * r * x​
# Avoid issues if f_prime is exactly zero (rare)​
if abs(f_prime)
> 1e-10:​
log_f_prime_sum += np.log(abs(f_prime))​
else:​
# Handle case of zero derivative if needed, e.g., skip or add large negative number​
# For logistic map at r=3.9, this is unlikely for typical x​
pass​
# Calculate next x​
x = r * x * (1 - x)​
​
N = n_total - n_skip​
lyapunov_exponent = log_f_prime_sum / N​
print(f"Estimated Lyapunov Exponent for r={r}: {lyapunov_exponent:.4f}")​

7.​ Expected Result: Running the code should yield a positive value for λ. For
r=3.9, the expected Lyapunov exponent is approximately λ≈0.4−0.5. A positive
Lyapunov exponent is a key indicator of chaotic behavior.
○​ Note: The exact value depends slightly on x0​(though it shouldn't for long N in
chaotic regime) and N. Using a larger N generally improves the estimate's
accuracy.

Exercise 9: Interpretation of Positive Lyapunov Exponent

●​ Problem: The calculation in Exercise 8 for the logistic map at r=3.9 yielded a
positive Lyapunov exponent (λ>0). Explain the significance of this result in terms
of the system's predictability and sensitivity to initial conditions.
●​ Worked Solution:
1.​ Definition Connection: The Lyapunov exponent λ measures the average
exponential rate of divergence or convergence of nearby trajectories in the
system's phase space. Specifically, if two initial conditions x0​and x0​+δ0​are
infinitesimally close, their separation δn​after n iterations evolves
approximately as ∣δn​∣≈∣δ0​∣eλn.
2.​ Positive Exponent Implication: A positive λ means that eλn grows
exponentially with n. Therefore, nearby initial conditions will, on average,
 separate exponentially fast as the system evolves.
3.​ Sensitivity to Initial Conditions: This exponential separation is the hallmark
of chaos, often referred to as the "butterfly effect." Even a minuscule
difference in the starting point x0​will lead to vastly different outcomes (xn​)
after a relatively small number of iterations.
4.​ Predictability: Consequently, long-term prediction becomes practically
impossible for a system with a positive Lyapunov exponent. Any tiny error in
measuring the initial state (which is unavoidable in practice) will be amplified
exponentially, rendering predictions inaccurate beyond a short time horizon.
The system's behavior is deterministic (following the map equation exactly),
but it is fundamentally unpredictable in the long run due to this sensitivity.
5.​ Contrast with Negative Exponent: If λ were negative, nearby trajectories
would converge exponentially (eλn→0), indicating stability and predictability.
The system would be insensitive to small changes in initial conditions in the
long term. A zero exponent typically indicates marginal stability or
quasi-periodicity.
○​ Thus, the positive Lyapunov exponent calculated for r=3.9 confirms the
chaotic nature observed visually in Exercise 7 and quantifies the rate at which
predictability is lost in this regime of the logistic map.

Subsection 1.3: Other Physical/Scientific Concepts

Exercise 10: Fourier Transform Application to Physical Signal
●​ Problem: A physical process generates a time-varying signal s(t) represented
over one period T=2π by s(t)=Asin(t)+0.5Asin(3t). Assume techniques like Fourier
analysis are used to analyze such signals. Calculate the complex Fourier series
coefficients cn​for this signal.
●​ Worked Solution:
1.​ Signal: s(t)=Asin(t)+0.5Asin(3t). The fundamental frequency is
ω0​=2π/T=2π/(2π)=1.
2.​ Complex Fourier Series: s(t)=∑n=−∞∞​cn​einω0​t=∑n=−∞∞​cn​eint.
3.​ Euler's Formula: Recall sin(x)=2ieix−e−ix​.
4.​ Rewrite Signal: Substitute Euler's formula into s(t):
s(t)=A(2ieit−e−it​)+0.5A(2iei3t−e−i3t​)
s(t)=2iA​eit−2iA​e−it+2i0.5A​ei3t−2i0.5A​e−i3t
s(t)=(−2iA​)eit+(2iA​)e−it+(−4iA​)ei3t+(4iA​)e−i3t
5.​ Identify Coefficients: Compare this expression with the complex Fourier
series form s(t)=∑cn​eint. We can directly read off the non-zero coefficients:
■​ For n=1: c1​=−iA/2
 ■​ For n=−1: c−1​=iA/2
■​ For n=3: c3​=−iA/4
■​ For n=−3: c−3​=iA/4
■​ For all other integer values of n (including n=0), cn​=0.
6.​ Result: The non-zero complex Fourier coefficients are c1​=−iA/2, c−1​=iA/2,
c3​=−iA/4, and c−3​=iA/4.
○​ Note: This demonstrates how a signal composed of simple sinusoidal
components can be represented in the frequency domain using Fourier
coefficients. The coefficients capture the amplitude and phase of each
frequency component present in the signal.

Exercise 11: Interpreting Frequency Spectrum of Physical Signal

●​ Problem: Based on the Fourier coefficients found in Exercise 10 for the signal
s(t)=Asin(t)+0.5Asin(3t), describe the frequency content of the signal. What does
the magnitude ∣cn​∣ represent? How does this relate to the original signal
composition?
●​ Worked Solution:
1.​ Non-zero Frequencies: The non-zero coefficients c1​,c−1​,c3​,c−3​indicate that
the signal s(t) contains frequency components only at nω0​=1⋅1=1 radian/sec
and nω0​=3⋅1=3 radians/sec (and their negative counterparts, which are
necessary for representing a real-valued signal).
2.​ Magnitude of Coefficients: The magnitude ∣cn​∣ represents the strength or
amplitude associated with the frequency component nω0​.
■​ ∣c1​∣=∣−iA/2∣=A/2.
■​ ∣c−1​∣=∣iA/2∣=A/2.
■​ ∣c3​∣=∣−iA/4∣=A/4.
■​ ∣c−3​∣=∣iA/4∣=A/4.
■​ ∣cn​∣=0 for all other n.
3.​ Relationship to Real Amplitude: For a real signal represented as a sum of
sines/cosines, the amplitude of the term cos(nω0​t) or sin(nω0​t) is related to
∣cn​∣ and ∣c−n​∣. Specifically, for a term An​sin(nω0​t), we have cn​=−iAn​/2 and
c−n​=iAn​/2, so ∣cn​∣=∣c−n​∣=An​/2.
4.​ Interpretation:
■​ The magnitudes ∣c1​∣=∣c−1​∣=A/2 correspond to the amplitude A of the
sin(t) term in the original signal (A1​=A).
■​ The magnitudes ∣c3​∣=∣c−3​∣=A/4 correspond to the amplitude 0.5A of the
sin(3t) term in the original signal (A3​=0.5A).
■​ The fact that all other ∣cn​∣ are zero confirms that the signal only contains
these two frequencies.
 ○​ Conclusion: The Fourier spectrum (plot of ∣cn​∣ vs nω0​) would show peaks
only at frequencies ω=1 and ω=3. The height of the peak at ω=1 would be
proportional to A/2, and the height of the peak at ω=3 would be proportional
to A/4. This spectrum provides a clear decomposition of the signal into its
constituent frequencies, which is a primary goal of Fourier analysis in signal
processing. Analyzing the frequency content can reveal underlying
periodicities or characteristic frequencies of the physical process generating
the signal.

Exercise 12: Simplified Navier-Stokes: Couette Flow

●​ Problem: Consider the steady, incompressible flow of a Newtonian fluid between
two infinite parallel plates separated by a distance h. The bottom plate (y=0) is
stationary, and the top plate (y=h) moves with a constant velocity U in the
x-direction. Assume the flow is purely in the x-direction, u=u(y), v=0, w=0, and
there is no pressure gradient in the x-direction (∂p/∂x=0). Simplify the
Navier-Stokes equations for this scenario and find the velocity profile u(y).
●​ Worked Solution:
1.​ Navier-Stokes Equations (Incompressible, Newtonian):
■​ Continuity: ∇⋅v=∂x∂u​+∂y∂v​+∂z∂w​=0
■​ x-momentum:
ρ(∂t∂u​+u∂x∂u​+v∂y∂u​+w∂z∂u​)=−∂x∂p​+μ(∂x2∂2u​+∂y2∂2u​+∂z2∂2u​)+ρgx​
■​ y-momentum:... (similar)
■​ z-momentum:... (similar) (where ρ is density, μ is dynamic viscosity, p is
pressure, g is gravity).
2.​ Apply Assumptions:
■​ Steady flow: ∂t∂​=0.
■​ v=0,w=0.
■​ u=u(y) only (implies ∂x∂u​=0, ∂z∂u​=0).
■​ ∂x∂p​=0.
■​ Neglect gravity or assume it only affects pressure hydrostatically (gx​=0).
3.​ Simplify Continuity: ∂x∂u​+∂y∂v​+∂z∂w​=0+0+0=0. (Satisfied).
4.​ Simplify x-momentum: ρ(0+u(0)+0∂y∂u​+0)=−0+μ(0+∂y2∂2u​+0)+0
0=μdy2d2u​(Since u only depends on y, partial derivatives become ordinary
derivatives).
5.​ Simplify y-momentum and z-momentum: These reduce to ∂y∂p​=ρgy​and
∂z∂p​=ρgz​, describing hydrostatic pressure variation, which doesn't affect the
velocity profile u(y).
6.​ Solve Velocity Equation: dy2d2u​=0.
■​ Integrate once: dydu​=C1​.
 ■​ Integrate again: u(y)=C1​y+C2​.
7.​ Apply Boundary Conditions:
■​ At y=0 (bottom plate): u(0)=0. ⟹C1​(0)+C2​=0⟹C2​=0.
■​ At y=h (top plate): u(h)=U. ⟹C1​h+C2​=U. Since C2​=0, C1​h=U⟹C1​=U/h.
8.​ Final Velocity Profile: Substitute C1​and C2​back: u(y)=hU​y
○​ Interpretation: This linear velocity profile is known as plane Couette flow. It
represents one of the simplest exact solutions to the Navier-Stokes equations,
illustrating shear-driven flow.

Exercise 13: Network Theory: Degree Distribution

●​ Problem: Consider a small network represented by the following adjacency list:
○​ Node A:
○​ Node B: [A, C, D]
○​ Node C:
○​ Node D:
○​ Node E: [D] Calculate the degree of each node and determine the degree
distribution P(k) (the probability that a randomly chosen node has degree k).
●​ Worked Solution:
1.​ Calculate Node Degrees: The degree of a node is the number of
connections it has.
■​ Degree(A) = 2 (connected to B, C)
■​ Degree(B) = 3 (connected to A, C, D)
■​ Degree(C) = 3 (connected to A, B, D)
■​ Degree(D) = 3 (connected to B, C, E)
■​ Degree(E) = 1 (connected to D)
2.​ List of Degrees: The degrees in the network are {2, 3, 3, 3, 1}.
3.​ Count Occurrences of Each Degree:
■​ Degree k=1: Occurs 1 time (Node E)
■​ Degree k=2: Occurs 1 time (Node A)
■​ Degree k=3: Occurs 3 times (Nodes B, C, D)
■​ Degree k=4, 5,...: Occur 0 times.
4.​ Calculate Degree Distribution P(k): There are N=5 nodes in total. P(k) is the
number of nodes with degree k divided by N.
■​ P(1)=1/5=0.2
■​ P(2)=1/5=0.2
■​ P(3)=3/5=0.6
■​ P(k)=0 for k≥4.
5.​ Result: The degree distribution is: P(1)=0.2, P(2)=0.2, P(3)=0.6.
○​ Interpretation: The degree distribution is a fundamental property of a
 network, characterizing the heterogeneity of connections. In larger networks
(e.g., social networks, internet), the shape of P(k) (e.g., power-law,
exponential) provides insights into the network's structure and formation
mechanisms.

Exercise 14: Entropy Calculation (Simple System)

●​ Problem: Consider a system of N=4 distinguishable particles that can each
occupy one of two energy levels: ϵ0​=0 and ϵ1​=ϵ. If the total energy of the system
is fixed at E=2ϵ, calculate the statistical entropy S=kB​lnΩ, where Ω is the number
of microstates corresponding to the macrostate E=2ϵ.
●​ Worked Solution:
1.​ Macrostate: The macrostate is defined by N=4 particles and total energy
E=2ϵ.
2.​ Microstates: A microstate specifies the energy level of each distinguishable
particle. To achieve a total energy of 2ϵ, exactly two particles must be in the
level ϵ1​=ϵ, and the other two must be in the level ϵ0​=0.
3.​ Counting Microstates (Ω): We need to choose which 2 of the 4
distinguishable particles are in the higher energy state ϵ. The number of ways
to do this is given by the binomial coefficient "4 choose 2":
Ω=(kN​)=(24​)=2!(4−2)!4!​=2!2!4!​=(2×1)(2×1)4×3×2×1​=424​=6
4.​ List Microstates (Optional): Let particles be 1, 2, 3, 4. States with energy 2ϵ:
■​ (1, 2 in ϵ1​; 3, 4 in ϵ0​)
■​ (1, 3 in ϵ1​; 2, 4 in ϵ0​)
■​ (1, 4 in ϵ1​; 2, 3 in ϵ0​)
■​ (2, 3 in ϵ1​; 1, 4 in ϵ0​)
■​ (2, 4 in ϵ1​; 1, 3 in ϵ0​)
■​ (3, 4 in ϵ1​; 1, 2 in ϵ0​) There are indeed 6 microstates.
5.​ Calculate Entropy: Use Boltzmann's entropy formula: S=kB​lnΩ. S=kB​ln(6)
○​ Result: The statistical entropy of the system with total energy E=2ϵ is
S=kB​ln(6). Entropy quantifies the number of microscopic arrangements
consistent with the macroscopic state.

Exercise 15: Boltzmann Distribution Application

●​ Problem: Consider a single particle that can exist in three energy states: E1​=0,
E2​=ϵ, and E3​=2ϵ. The particle is in thermal equilibrium with a reservoir at
temperature T. Calculate the probability Pi​of finding the particle in each energy
state i=1,2,3. What happens to the probabilities as T→0 and T→∞?
●​ Worked Solution:
1.​ Boltzmann Distribution: The probability Pi​of finding the particle in state i
 with energy Ei​is given by: Pi​=Ze−Ei​/(kB​T)​where Z is the partition function.
2.​ Partition Function (Z): The partition function is the sum of the Boltzmann
factors over all states: Z=j=1∑3​e−Ej​/(kB​T)=e−E1​/(kB​T)+e−E2​/(kB​T)+e−E3​/(kB​T)
Z=e−0/(kB​T)+e−ϵ/(kB​T)+e−2ϵ/(kB​T) Z=1+e−βϵ+e−2βϵ where β=1/(kB​T).
3.​ Calculate Probabilities:
■​ P1​=Ze−E1​β​=Ze0​=1+e−βϵ+e−2βϵ1​
■​ P2​=Ze−E2​β​=1+e−βϵ+e−2βϵe−βϵ​
■​ P3​=Ze−E3​β​=1+e−βϵ+e−2βϵe−2βϵ​
4.​ Limit T → 0 (Low Temperature):
■​ As T→0, β=1/(kB​T)→∞.
■​ e−βϵ→0 and e−2βϵ→0 (assuming ϵ>0).
■​ Z→1+0+0=1.
■​ P1​→1/1=1.
■​ P2​→0/1=0.
■​ P3​→0/1=0.
■​ Interpretation: At absolute zero, the particle is certain to be found in the
lowest energy state (ground state).
5.​ Limit T → ∞ (High Temperature):
■​ As T→∞, β=1/(kB​T)→0.
■​ e−βϵ→e0=1.
■​ e−2βϵ→e0=1.
■​ Z→1+1+1=3.
■​ P1​→1/3.
■​ P2​→1/3.
■​ P3​→1/3.
■​ Interpretation: At very high temperatures, thermal energy vastly exceeds
the energy level differences. All energy states become equally probable.

Part 2: Exercises in Stock Price Prediction

This section focuses on quantitative models and techniques commonly applied to
financial markets, particularly stock price analysis, including stochastic processes,
time series modeling, and concepts potentially borrowed from physics.

Subsection 2.1: Stochastic Processes and Models

Exercise 16: GBM: Expected Value of Stock Price
●​ Problem: Assume a stock price St​follows Geometric Brownian Motion (GBM) as
described by the stochastic differential equation (SDE): dSt​=μSt​dt+σSt​dWt​,
where μ is the expected return (drift), σ is the volatility, and Wt​is a standard
 Wiener process. Given the initial price S0​, find the expected stock price at time t,
E.
●​ Worked Solution:
1.​ GBM Solution: The solution to the GBM SDE is St​=S0​exp((μ−2σ2​)t+σWt​).
2.​ Expectation: We need to compute $E = E\left$.
3.​ Properties: S0​is a constant. The term (μ−2σ2​)t is deterministic. Wt​is a
normally distributed random variable with E[Wt​]=0 and Var(Wt​)=t. Therefore,
σWt​is normally distributed with mean E[σWt​]=σE[Wt​]=0 and variance
Var(σWt​)=σ2Var(Wt​)=σ2t.
4.​ Lognormal Property: Let X=σWt​. Then X∼N(0,σ2t). We need the expectation
of eX. For a normally distributed random variable X∼N(m,v), the expectation of
its exponential is E[eX]=em+v/2.
5.​ Apply Property: In our case, X=σWt​, so m=0 and v=σ2t. Therefore,
E[eσWt​]=e0+(σ2t)/2=eσ2t/2.
6.​ Calculate Expectation: E=S0​E[exp((μ−2σ2​)t)exp(σWt​)]
E=S0​exp((μ−2σ2​)t)E[exp(σWt​)] E=S0​exp(μt−2σ2​t)eσ2t/2 E=S0​eμt−2σ2​t+2σ2​t
E=S0​eμt
○​ Result: The expected stock price under GBM grows exponentially at the rate
of the drift parameter μ: E=S0​eμt.

Exercise 17: GBM: Variance of Stock Price

●​ Problem: For the same GBM process St​=S0​exp((μ−2σ2​)t+σWt​), find the variance
of the stock price at time t, Var(St​).
●​ Worked Solution:
1.​ Variance Formula: Var(St​)=E−(E)2. We already know E=S0​eμt from Exercise
16.
2.​ Calculate E: $$S_t^2 = \left^2$$ St2​=S02​exp(2(μ−2σ2​)t+2σWt​) $$E = E\left$$
E=S02​exp((2μ−σ2)t)E[exp(2σWt​)]
3.​ Expectation of Exponential: Let Y=2σWt​. Since Wt​∼N(0,t),
Y∼N(E[2σWt​],Var(2σWt​)).
■​ E[Y]=2σE[Wt​]=0.
■​ Var(Y)=(2σ)2Var(Wt​)=4σ2t. So, Y∼N(0,4σ2t).
4.​ Using the lognormal property E[eY]=em+v/2 with m=0 and v=4σ2t:
E[exp(2σWt​)]=e0+(4σ2t)/2=e2σ2t
5.​ Substitute back into E: E=S02​exp((2μ−σ2)t)e2σ2t E=S02​e2μt−σ2t+2σ2t
E=S02​e2μt+σ2t
6.​ Calculate Variance: Var(St​)=E−(E)2 Var(St​)=S02​e2μt+σ2t−(S0​eμt)2
Var(St​)=S02​e2μteσ2t−S02​e2μt Var(St​)=S02​e2μt(eσ2t−1)
○​ Result: The variance of the stock price under GBM is
 Var(St​)=S02​e2μt(eσ2t−1). Note that the variance grows over time due to both
the drift and the volatility term.

Exercise 18: GBM: Expected Value & Variance of Log-Returns

●​ Problem: Define the continuously compounded (log) return over a period t as
Rt​=ln(St​/S0​). For a stock price St​following GBM, find the expected log-return E
and the variance of the log-return Var(Rt​). What distribution does the log-return
follow?
●​ Worked Solution:
1.​ GBM Solution: St​=S0​exp((μ−2σ2​)t+σWt​).
2.​ Calculate Log-Return Rt​: Rt​=ln(S0​St​​)=ln(exp((μ−2σ2​)t+σWt​))
Rt​=(μ−2σ2​)t+σWt​
3.​ Calculate Expected Log-Return E: E=E[(μ−2σ2​)t+σWt​] Since (μ−2σ2​)t is
deterministic and E[Wt​]=0: E=(μ−2σ2​)t+σE[Wt​] E=(μ−2σ2​)t
4.​ Calculate Variance of Log-Return Var(Rt​): Var(Rt​)=Var[(μ−2σ2​)t+σWt​] Since
adding a constant does not change variance, Var(X+c)=Var(X):
Var(Rt​)=Var(σWt​) Using Var(aX)=a2Var(X) and Var(Wt​)=t: Var(Rt​)=σ2Var(Wt​)
Var(Rt​)=σ2t
5.​ Distribution of Log-Return: The log-return Rt​=(μ−2σ2​)t+σWt​is a constant
term plus a normally distributed term (σWt​∼N(0,σ2t)). A linear transformation
of a normal random variable is also normal. Therefore, Rt​follows a normal
distribution with mean (μ−2σ2​)t and variance σ2t. Rt​∼N((μ−2σ2​)t,σ2t)
○​ Result: The log-return Rt​=ln(St​/S0​) under GBM is normally distributed with
mean (μ−2σ2​)t and variance σ2t. This normality of log-returns is a key
prediction of the GBM model.

Exercise 19: GBM Path Simulation (Euler-Maruyama)

●​ Problem: Simulate one sample path of a stock price following GBM
dSt​=μSt​dt+σSt​dWt​using the Euler-Maruyama method. Use parameters S0​=100,
μ=0.10 (10% annual drift), σ=0.20 (20% annual volatility). Simulate daily prices for
one year (T=1, number of steps N=252). Use a time step Δt=T/N. Recall that dWt​
can be approximated by ZΔt​, where Z∼N(0,1).
●​ Worked Solution:
1.​ Euler-Maruyama Scheme: The discrete approximation of the SDE
dSt​=a(St​,t)dt+b(St​,t)dWt​is Si+1​=Si​+a(Si​,ti​)Δt+b(Si​,ti​)ΔWi​.
2.​ Apply to GBM: Here a(St​,t)=μSt​and b(St​,t)=σSt​. ΔWi​≈Zi​Δt​, where Zi​are
independent standard normal random numbers. Si+1​=Si​+μSi​Δt+σSi​Zi​Δt​
Si+1​=Si​(1+μΔt+σZi​Δt​)
3.​ Parameters:
 ■​ S0​=100
■​ μ=0.10
■​ σ=0.20
■​ T=1 year
■​ N=252 steps (trading days)
■​ Δt=T/N=1/252
4.​ Simulation Steps:
■​ Initialize S=S0​=100.
■​ For i from 0 to N−1:
■​ Generate a random number Zi​from N(0,1).
■​ Calculate S[i+1]=S[i]∗(1+μ∗Δt+σ∗Zi​∗Δt​).
■​ The array S,S,...,S[N] represents the simulated daily stock prices for one
year.
5.​ Conceptual Code (Python):​
Python​
import numpy as np​
import matplotlib.pyplot as plt​
​
S0 = 100.0​
mu = 0.10​
sigma = 0.20​
T = 1.0​
N = 252​
dt = T / N​
​
# Generate random numbers​
Z = np.random.standard_normal(N)​
# Or use: Z = np.random.normal(0.0, 1.0, N)​
​
S = np.zeros(N + 1)​
S = S0​
​
# Simulation loop​
for i in range(N):​
S[i+1] = S[i] * (1 + mu * dt + sigma * Z[i] * np.sqrt(dt))​
​
# Plotting​
time = np.linspace(0, T, N + 1)​
plt.plot(time, S)​
plt.xlabel("Time (Years)")​
 plt.ylabel("Stock Price")​
plt.title("GBM Path Simulation (Euler-Maruyama)")​
plt.grid(True)​
plt.show()​

○​ Result: Running the code will produce a plot showing a jagged path starting
at 100, representing one possible evolution of the stock price under the GBM
assumptions. Repeating the simulation will yield different paths due to the
random nature of Zi​. This method is fundamental for Monte Carlo simulations
in finance.

Exercise 20: Ito's Lemma Application (log(S_t) for GBM)

●​ Problem: Let St​follow the GBM process dSt​=μSt​dt+σSt​dWt​. Use Ito's Lemma to
find the SDE for the process f(St​)=ln(St​).
●​ Worked Solution:
1.​ Ito's Lemma: For a process Xt​following dXt​=a(Xt​,t)dt+b(Xt​,t)dWt​, and a
twice-differentiable function f(Xt​,t), the process Yt​=f(Xt​,t) follows the SDE:
dYt​=(∂t∂f​+a(Xt​,t)∂x∂f​+21​b(Xt​,t)2∂x2∂2f​)dt+b(Xt​,t)∂x∂f​dWt​
2.​ Identify Terms:
■​ Our process is St​, so Xt​=St​.
■​ The drift is a(St​,t)=μSt​.
■​ The diffusion is b(St​,t)=σSt​.
■​ Our function is f(St​)=ln(St​). It does not explicitly depend on t.
3.​ Calculate Derivatives of f:
■​ ∂t∂f​=0
■​ ∂St​∂f​=St​1​
■​ ∂St2​∂2f​=−St2​1​
4.​ Substitute into Ito's Lemma: Let Yt​=ln(St​).
d(lnSt​)=(0+(μSt​)(St​1​)+21​(σSt​)2(−St2​1​))dt+(σSt​)(St​1​)dWt​
5.​ Simplify: d(lnSt​)=(μ+21​σ2St2​(−St2​1​))dt+σdWt​d(lnSt​)=(μ−21​σ2)dt+σdWt​
○​ Result: The SDE for the log-price process ln(St​) is d(lnSt​)=(μ−2σ2​)dt+σdWt​.
This confirms that the log-price follows a generalized Wiener process
(arithmetic Brownian motion with drift). Integrating this SDE from 0 to t yields
ln(St​)−ln(S0​)=(μ−2σ2​)t+σ(Wt​−W0​). Assuming W0​=0, we get
ln(St​/S0​)=(μ−2σ2​)t+σWt​, which is exactly the log-return Rt​found in Exercise
18, consistent with the solution form of GBM. This exercise demonstrates the
power of Ito's Lemma for deriving the dynamics of functions of stochastic
processes.
Exercise 21: Discussion: GBM Limitations (Fat Tails)
●​ Problem: The GBM model predicts that log-returns Rt​=ln(St​/S0​) are normally
distributed (Exercise 18). However, empirical studies of financial asset returns
often show "fat tails" (leptokurtosis). Explain what "fat tails" means and why this
empirical observation poses a challenge to the basic GBM model.
●​ Worked Solution:
1.​ Normal Distribution Tails: The normal distribution has tails that decay very
rapidly (exponentially with the square of the deviation from the mean). This
means that extreme events (very large positive or negative returns) are
predicted to be extremely rare.
2.​ Fat Tails (Leptokurtosis): "Fat tails" refers to the property of a probability
distribution where the tails decay more slowly than the tails of a normal
distribution (e.g., like a power law). This implies that extreme events are
significantly more likely to occur in reality than predicted by a normal
distribution with the same mean and variance. Leptokurtosis is a statistical
measure (kurtosis > 3) indicating peakedness and heavier tails compared to a
normal distribution.
3.​ Empirical Observation: Numerous studies analyzing historical stock returns,
currency exchange rates, and other financial time series consistently find that
the distribution of their log-returns exhibits fat tails. Large daily or weekly
price jumps (both up and down) occur much more frequently than would be
expected under a normal distribution assumption.
4.​ Challenge to GBM: Since GBM inherently predicts normally distributed
log-returns, the empirical evidence of fat tails directly contradicts this
prediction. This suggests that the standard GBM model, while mathematically
convenient and foundational, fails to capture a crucial aspect of real-world
asset price dynamics – the occurrence of unexpectedly large price
movements.
5.​ Implications: Using a pure GBM model for applications like risk management
(e.g., Value-at-Risk calculations) or option pricing can lead to a significant
underestimation of the probability of large losses (or gains). The model might
suggest that a 5-standard-deviation move is practically impossible, while
empirical data shows such events happen, albeit infrequently. This
discrepancy highlights the need for more advanced models that can
accommodate fat tails, such as models with stochastic volatility (like GARCH
or Heston) or jump-diffusion processes. The idealized nature of GBM, while
useful as a starting point, requires refinement to better match observed
market behavior.
Subsection 2.2: Time Series Analysis and Volatility Modeling
Exercise 22: GARCH(1,1) Volatility Forecasting
●​ Problem: A GARCH(1,1) model is often used to model the time-varying volatility of
financial returns. The model equations are:
○​ Return: rt​=σt​ϵt​, where ϵt​∼N(0,1) (or other distribution with mean 0, variance 1).
○​ Conditional Variance: σt2​=ω+αrt−12​+βσt−12​. Suppose the estimated
parameters for a daily stock return series are ω=0.000002, α=0.09, and
β=0.90. If the variance estimate for yesterday (day t−1) was σt−12​=0.000150
(corresponding to σt−1​≈1.22% daily vol) and yesterday's return was
rt−1​=−0.025 (-2.5%), forecast the variance σt2​for today (day t).
●​ Worked Solution:
1.​ GARCH(1,1) Variance Equation: σt2​=ω+αrt−12​+βσt−12​.
2.​ Given Information:
■​ ω=0.000002
■​ α=0.09
■​ β=0.90
■​ rt−1​=−0.025
■​ σt−12​=0.000150
3.​ Calculate rt−12​: rt−12​=(−0.025)2=0.000625.
4.​ Substitute Values into the Equation:
σt2​=0.000002+(0.09×0.000625)+(0.90×0.000150)
σt2​=0.000002+0.00005625+0.000135 σt2​=0.00019325
○​ Result: The forecast for today's variance is σt2​=0.00019325.
○​ Interpretation: The forecasted daily volatility is σt​=0.00019325​≈0.0139 or
1.39%. Yesterday's large negative return (rt−1​=−2.5%) increased the variance
forecast for today through the αrt−12​term, demonstrating how GARCH
models capture volatility clustering (large price changes tend to be followed
by large price changes).

Exercise 23: GARCH(1,1) Unconditional Variance

●​ Problem: For the GARCH(1,1) process σt2​=ω+αrt−12​+βσt−12​, where rt−1​=σt−1​ϵt−1​
with E[ϵt−12​]=1, find the long-run average or unconditional variance,
Var(rt​)=E[rt2​]=E[σt2​], assuming the process is stationary (α+β<1). Use the
parameters from Exercise 22: ω=0.000002, α=0.09, β=0.90.
●​ Worked Solution:
1.​ Stationarity Condition: Check if α+β<1. Here, 0.09+0.90=0.99<1. The
process is stationary.
2.​ Unconditional Variance Definition: If the process is stationary, the
 unconditional variance σ2=E[σt2​] is constant over time. Also,
E[σt2​]=E[σt−12​]=σ2. Furthermore,
E[rt−12​]=E[E[rt−12​∣Ft−2​]]=E[E[σt−12​ϵt−12​∣Ft−2​]]=E[σt−12​E[ϵt−12​]]=E[σt−12​](1)
=σ2.
3.​ Take Expectation of GARCH Equation: Take the unconditional expectation
of both sides of the variance equation: E[σt2​]=E[ω+αrt−12​+βσt−12​]
4.​ Use Linearity of Expectation: E[σt2​]=E[ω]+E[αrt−12​]+E[βσt−12​]
E[σt2​]=ω+αE[rt−12​]+βE[σt−12​]
5.​ Substitute Unconditional Variance σ2: σ2=ω+ασ2+βσ2
6.​ Solve for σ2: σ2−ασ2−βσ2=ω σ2(1−α−β)=ω σ2=1−α−βω​
7.​ Substitute Parameter Values:
σ2=1−0.09−0.900.000002​=1−0.990.000002​=0.010.000002​σ2=0.0002
○​ Result: The unconditional variance of the process is σ2=0.0002.
○​ Interpretation: This represents the long-run average level around which the
conditional variance σt2​fluctuates. The corresponding long-run average daily
volatility is σ=0.0002​≈0.0141 or 1.41%. The GARCH forecast from Exercise 22
(1.39%) was close to this long-run average but slightly higher due to the
recent large return.

Exercise 24: Interpretation of GARCH Parameters (Persistence)

●​ Problem: In the GARCH(1,1) model σt2​=ω+αrt−12​+βσt−12​, the sum α+β is often
referred to as the "persistence" parameter. Using the parameters from Exercise 22
(α=0.09,β=0.90), calculate the persistence and explain what a high value (close
to 1) implies about the behavior of volatility shocks.
●​ Worked Solution:
1.​ Calculate Persistence: Persistence = α+β=0.09+0.90=0.99.
2.​ Interpretation:
■​ The GARCH equation shows that today's variance σt2​is influenced by
yesterday's squared return rt−12​(the "shock" or news component,
weighted by α) and yesterday's variance σt−12​(the persistent component,
weighted by β).
■​ The sum α+β determines how quickly the conditional variance reverts to
its long-run mean σ2=ω/(1−α−β) after a shock.
■​ A value of α+β close to 1 indicates high persistence. This means that
shocks to volatility (periods of unusually high or low variance) tend to die
out slowly.
■​ If there is a large shock (large rt−12​), it increases σt2​. Because β is large
(0.90 in this case), this higher variance level significantly influences σt+12​,
which in turn influences σt+22​, and so on. The effect of the initial shock
 decays at a rate related to (α+β)k for k steps ahead.
■​ With α+β=0.99, the decay is very slow. A period of high volatility will tend
to be followed by further periods of high volatility, and similarly for low
volatility. This mathematical persistence directly models the empirical
phenomenon known as "volatility clustering" often observed in financial
markets – the tendency for volatile periods and calm periods to group
together.
■​ If α+β were small, shocks would dissipate quickly, and volatility would
rapidly revert to its long-run mean. If α+β≥1, the model implies
non-stationarity (Integrated GARCH or IGARCH), where shocks have a
permanent effect on the variance level.

Exercise 25: Stationarity Testing Concept (e.g., ADF)

●​ Problem: Many time series models (like ARMA) assume the underlying process is
stationary. Briefly explain the concept of stationarity (specifically, weak
stationarity). Why is it important? Mention a common test used to check for
non-stationarity (e.g., the Augmented Dickey-Fuller test described conceptually
in). What is the typical null hypothesis of such a test?
●​ Worked Solution:
1.​ Weak Stationarity: A time series process {Yt​} is weakly stationary (or
covariance stationary) if its statistical properties do not change over time.
Specifically, it requires:
■​ Constant Mean: E[Yt​]=μ (independent of t).
■​ Constant Variance: Var(Yt​)=σ2 (independent of t).
■​ Constant Autocovariance: Cov(Yt​,Yt−k​)=γk​(depends only on the lag k, not
on time t).
2.​ Importance: Stationarity is crucial because:
■​ Model Validity: Many standard time series models (e.g., ARMA) are
developed under the assumption of stationarity. Applying them to
non-stationary data can lead to spurious regressions and unreliable
forecasts.
■​ Predictability: In a stationary process, patterns observed in the past (like
autocorrelations) are likely to persist, making forecasting meaningful. In a
non-stationary process (like a random walk), the mean or variance may
drift unpredictably.
■​ Inference: Statistical inference (hypothesis testing, confidence intervals)
based on models assuming stationarity can be invalid if the data is
non-stationary.
3.​ Common Test (ADF): The Augmented Dickey-Fuller (ADF) test is a widely
 used statistical test for the presence of a unit root, which is a common cause
of non-stationarity in time series data. A unit root implies that shocks to the
series have permanent effects, and the variance grows over time. The ADF
test extends the basic Dickey-Fuller test to handle more complex models with
higher-order autoregressive terms.
4.​ Null Hypothesis: The null hypothesis (H0​) of the ADF test (and the basic
Dickey-Fuller test) is that the time series has a unit root (i.e., it is
non-stationary).
5.​ Alternative Hypothesis: The alternative hypothesis (H1​) is that the time
series does not have a unit root (i.e., it is stationary, possibly after removing a
deterministic trend).
6.​ Procedure Overview: The test involves estimating a regression model and
examining the coefficient of the lagged level term. The calculated test statistic
is compared to critical values from the Dickey-Fuller distribution. If the test
statistic is more negative than the critical value, the null hypothesis of a unit
root is rejected, providing evidence for stationarity. If non-stationarity is
detected, transformations like differencing are often applied to make the
series stationary before modeling.

Subsection 2.3: Econophysics and Advanced Models

Exercise 26: Black-Scholes PDE Derivation Step (Heat Eq. Analogy)
●​ Problem: The derivation of the Black-Scholes Partial Differential Equation (PDE)
for option pricing involves constructing a risk-free portfolio and applying Ito's
Lemma. A key step involves transforming the resulting PDE into a form resembling
the heat equation. Let the Black-Scholes PDE be:
∂t∂V​+rS∂S∂V​+21​σ2S2∂S2∂2V​−rV=0 where V(S,t) is the option price, S is the stock
price, t is time, r is the risk-free rate, and σ is the volatility. Apply the following
change of variables, motivated by the heat equation analogy:
○​ S=ex (or x=lnS)
○​ t=T−σ22τ​(where T is expiry time, τ is time variable for heat equation)
○​ V(S,t)=eax+bτu(x,τ) (for some constants a,b to be determined) Show that by
choosing appropriate a and b, the PDE for u(x,τ) can be reduced to the
standard heat equation ∂τ∂u​=∂x2∂2u​.
●​ Worked Solution:
1.​ Calculate Derivatives: We need to express the derivatives of V with respect
to S and t in terms of derivatives of u with respect to x and τ. Use the chain
rule.
■​ ∂S∂V​=∂x∂V​∂S∂x​=∂x∂V​S1​.
 ■​ ∂S2∂2V​=∂S∂​(S1​∂x∂V​)=−S21​∂x∂V​+S1​∂S∂​(∂x∂V​)=−S21​∂x∂V​+S1​(∂x2∂2V​∂S∂x​
)=S21​(∂x2∂2V​−∂x∂V​).
■​ ∂t∂V​=∂τ∂V​∂t∂τ​. From t=T−σ22τ​, we have dτdt​=−σ22​, so ∂t∂τ​=−2σ2​. Thus
∂t∂V​=−2σ2​∂τ∂V​.
2.​ Express V derivatives in terms of u derivatives: Let V=eax+bτu.
■​ ∂x∂V​=aeax+bτu+eax+bτ∂x∂u​=eax+bτ(au+∂x∂u​).
■​ ∂x2∂2V​=aeax+bτ(au+∂x∂u​)+eax+bτ(a∂x∂u​+∂x2∂2u​)=eax+bτ(a2u+2a∂x∂u​+
∂x2∂2u​).
■​ ∂τ∂V​=beax+bτu+eax+bτ∂τ∂u​=eax+bτ(bu+∂τ∂u​).
3.​ Substitute into Black-Scholes PDE: Substitute the expressions for
derivatives of V into the PDE. Note S=ex.
−2σ2​∂τ∂V​+rS(S1​∂x∂V​)+21​σ2S2(S21​(∂x2∂2V​−∂x∂V​))−rV=0
−2σ2​∂τ∂V​+r∂x∂V​+21​σ2(∂x2∂2V​−∂x∂V​)−rV=0
−2σ2​∂τ∂V​+(r−2σ2​)∂x∂V​+21​σ2∂x2∂2V​−rV=0
4.​ Substitute V=eax+bτu and its derivatives: Divide the entire equation by
eax+bτ. −2σ2​(bu+∂τ∂u​)+(r−2σ2​)(au+∂x∂u​)+21​σ2(a2u+2a∂x∂u​+∂x2∂2u​)−r(u)=0
5.​ Group terms by derivatives of u: $$ \left(-\frac{\sigma^2}{2}\right)
\frac{\partial u}{\partial \tau} + \left( (r - \frac{\sigma^2}{2}) +
\frac{1}{2}\sigma^2 (2a) \right) \frac{\partial u}{\partial x} +
\left(\frac{1}{2}\sigma^2\right) \frac{\partial^2 u}{\partial x^2} $$ $$ + \left(
-\frac{\sigma^2}{2} b + (r - \frac{\sigma^2}{2}) a + \frac{1}{2}\sigma^2 a^2 - r
\right) u = 0 $$
6.​ Simplify coefficients: $$ -\frac{\sigma^2}{2} \frac{\partial u}{\partial \tau} +
\left( r - \frac{\sigma^2}{2} + a\sigma^2 \right) \frac{\partial u}{\partial x} +
\frac{\sigma^2}{2} \frac{\partial^2 u}{\partial x^2} $$ $$ + \left(
-\frac{\sigma^2}{2} b + a(r - \frac{\sigma^2}{2}) + \frac{1}{2}\sigma^2 a^2 - r
\right) u = 0 $$
7.​ Choose a and b to simplify: We want the equation to become ∂τ∂u​=∂x2∂2u​.
■​ Divide the whole equation by 2σ2​: $$ -\frac{\partial u}{\partial \tau} +
\frac{2}{\sigma^2}\left( r - \frac{\sigma^2}{2} + a\sigma^2 \right)
\frac{\partial u}{\partial x} + \frac{\partial^2 u}{\partial x^2} $$ $$ +
\frac{2}{\sigma^2}\left( -\frac{\sigma^2}{2} b + a(r - \frac{\sigma^2}{2}) +
\frac{1}{2}\sigma^2 a^2 - r \right) u = 0 $$
■​ To eliminate the ∂x∂u​term, set its coefficient to zero:
r−2σ2​+aσ2=0⟹aσ2=2σ2​−r⟹a=21​−σ2r​.
■​ To eliminate the u term, set its coefficient to zero:
−2σ2​b+a(r−2σ2​)+21​σ2a2−r=0. Substitute a(r−2σ2​)=a(−aσ2)=−a2σ2.
−2σ2​b−a2σ2+21​σ2a2−r=0 −2σ2​b−21​σ2a2−r=0 b=−a2−σ22r​. Substitute
a=21​−σ2r​: a2=(21​−σ2r​)2=41​−σ2r​+σ4r2​.
 b=−(41​−σ2r​+σ4r2​)−σ22r​=−41​−σ2r​−σ4r2​. Let k=r/(σ2/2)=2r/σ2. Then
a=(1−k)/2. b=−(21−k​)2−k=−41−2k+k2​−k=−41+2k+k2​=−(21+k​)2. a=21​−σ2r​,
b=−(21​+σ2r​)2. Let's recheck the coefficient of u. The coefficient of u is:
−2σ2​b+a(r−2σ2​)+21​σ2a2−r. Let's try setting a and b to make the equation
∂τ∂u​=∂x2∂2u​. Comparing −∂τ∂u​+⋯+∂x2∂2u​+⋯=0 with ∂τ∂u​=∂x2∂2u​. We
need the coefficient of ∂x∂u​to be zero, and the coefficient of u to be zero.
Coefficient of ∂x∂u​: σ22​(r−2σ2​+aσ2)=0⟹a=21​−σ2r​. (Correct) Coefficient
of u: σ22​(−2σ2​b+a(r−2σ2​)+21​σ2a2−r)=0. −b+σ22a​(r−2σ2​)+a2−σ22r​=0.
Substitute a=21​−σ2r​: σ22a​=σ21​−σ42r​. r−2σ2​.
σ22a​(r−2σ2​)=(σ21​−σ42r​)(r−2σ2​)=σ2r​−21​−σ42r2​+σ2r​=σ22r​−21​−σ42r2​.
a2=(21​−σ2r​)2=41​−σ2r​+σ4r2​. Coefficient of u (multiplied by σ2/2):
−2σ2​b+a(r−2σ2​)+21​σ2a2−r=0. −2σ2​b+(21​−σ2r​)(r−2σ2​)+21​σ2(21​−σ2r​)2−r=0.
−2σ2​b+(2r​−4σ2​−σ2r2​+2r​)+21​σ2(41​−σ2r​+σ4r2​)−r=0.
−2σ2​b+(r−4σ2​−σ2r2​)+(8σ2​−2r​+2σ2r2​)−r=0. −2σ2​b−8σ2​−2r​−2σ2r2​=0.
2σ2​b=−8σ2​−2r​−2σ2r2​. b=−41​−σ2r​−σ4r2​=−(21​+σ2r​)2.
8.​ Final Equation: With these choices of a=21​−σ2r​and b=−(21​+σ2r​)2, the PDE
simplifies to: −∂τ∂u​+0⋅∂x∂u​+∂x2∂2u​+0⋅u=0 ∂τ∂u​=∂x2∂2u​
○​ Result: The transformation successfully converts the Black-Scholes PDE into
the standard 1D heat equation for u(x,τ). This allows the use of known solution
techniques for the heat equation (like Fourier transforms or Green's functions)
to find the solution for u, which can then be transformed back to find the
option price V(S,t). This analogy provides both a solution pathway and
conceptual link between financial modeling and physics.

Exercise 27: Critique of Black-Scholes Heat Equation Analogy

●​ Problem: The transformation in Exercise 26 highlights an analogy between the
Black-Scholes model and the heat equation. However, the Black-Scholes model
relies on several assumptions, including that the underlying asset follows GBM,
implying normally distributed log-returns. Considering the empirical evidence of
fat tails in financial returns, discuss the limitations or potential weaknesses of
relying heavily on this physics-based analogy.
●​ Worked Solution:
1.​ Foundation of the Analogy: The successful transformation relies entirely on
the validity of the Black-Scholes PDE, which itself is derived assuming the
underlying asset St​follows GBM: dSt​=μSt​dt+σSt​dWt​. A key consequence of
GBM is that log-returns ln(St​/S0​) are normally distributed.
2.​ Empirical Contradiction: As noted, real-world financial asset returns
frequently exhibit "fat tails" (leptokurtosis) and volatility clustering, meaning
extreme events are more common than predicted by the normal distribution
 inherent in GBM, and volatility is not constant but changes over time.
3.​ Weakness 1: Mismatch in Randomness: The heat equation typically models
diffusion driven by Brownian motion (many small, independent random
impacts), naturally leading to Gaussian distributions. The Black-Scholes
model inherits this via the dWt​term. If the actual driving "randomness" in
financial markets includes occasional large jumps or time-varying intensity
(stochastic volatility), which are needed to explain fat tails, then the
fundamental process being modeled differs significantly from standard heat
diffusion. The analogy breaks down at the level of the underlying stochastic
process.
4.​ Weakness 2: Constant Parameters: The transformation and the resulting
heat equation assume constant parameters (r, σ). While r might be relatively
stable in the short term, the assumption of constant volatility σ is particularly
problematic. Empirical evidence strongly suggests volatility is time-varying
(stochastic volatility). A heat equation with constant diffusivity α cannot
capture phenomena like volatility clustering. While Black-Scholes can be
viewed as pricing under an "implied" constant volatility, this differs from the
actual dynamics.
5.​ Weakness 3: Risk Implications: The normal distribution assumption
embedded in the analogy leads to underestimation of tail risk (the probability
of large losses). Models based purely on this analogy might provide
inaccurate risk assessments or option prices, especially for options sensitive
to extreme market moves (e.g., deep out-of-the-money options).
○​ Conclusion: While the mathematical transformation to the heat equation is
elegant and provides a powerful solution technique for the idealized
Black-Scholes world, its direct physical analogy can be misleading. The
empirical deviations from GBM, particularly fat tails and stochastic volatility,
suggest that the underlying processes in finance are more complex than
simple heat diffusion. Relying solely on the analogy without acknowledging
these differences risks mischaracterizing market behavior and
underestimating risk. More advanced models (e.g., jump-diffusions, stochastic
volatility models, GARCH) attempt to address these empirical realities, moving
beyond the limitations of the direct heat equation parallel.

Exercise 28: ABM Simulation Analysis: Strategy Switching Frequency

●​ Problem: An Agent-Based Model (ABM) simulates a financial market with agents
choosing between different trading strategies (e.g., fundamental vs. technical)
based on performance. Simulation output data provides a time series of the
fraction of agents using Strategy A, denoted fA​(t). Suppose over a period of 1000
 time steps, the total number of times agents switched from Strategy B to Strategy
A was 1500, and the total number of times agents switched from Strategy A to
Strategy B was 1450. If the total number of agents N is 500, calculate the average
switching frequency per agent per time step.
●​ Worked Solution:
1.​ Total Switches: The total number of strategy switches observed in the
simulation is the sum of switches in both directions: Total Switches =
(Switches B to A) + (Switches A to B) = 1500 + 1450 = 2950.
2.​ Simulation Duration: The simulation ran for T=1000 time steps.
3.​ Number of Agents: There are N=500 agents.
4.​ Average Switches per Time Step: The average number of switches
occurring across the entire population in a single time step is: Average
Switches / Step = Total Switches / T = 2950 / 1000 = 2.95 switches per step.
5.​ Average Switching Frequency per Agent: To find the average frequency
per agent, divide the average switches per step by the number of agents:
Average Frequency per Agent = (Average Switches / Step) / N = 2.95 / 500 =
0.0059 switches per agent per time step.
○​ Result: The average switching frequency is 0.0059 switches per agent per
time step.
○​ Interpretation: This metric quantifies the typical rate at which individual
agents reconsider and potentially change their trading strategy within the
model. A higher frequency might indicate a more reactive or less stable
market environment within the simulation, where agents quickly adapt to
perceived changes in strategy performance. Comparing this frequency under
different parameter settings (e.g., different performance evaluation rules) can
provide insights into the model's dynamics.

Exercise 29: ABM Simulation Analysis: Identifying Market Stress

●​ Problem: The same ABM simulation also outputs the market price P(t) and a
measure of market volatility V(t). The model description suggests that periods of
"market stress" are characterized by simultaneously high volatility
(V(t)>Vthreshold​) and a rapid decline in the price (e.g., P(t)−P(t−1)<−ΔPthreshold​).
Given thresholds Vthreshold​=0.05 and ΔPthreshold​=1.5, analyze the following
short segment of simulation output and identify the time steps corresponding to
market stress.

Time (t) Price P(t) Volatility V(t)

101 98.5 0.03

102 97.0 0.04

103 94.5 0.06

104 95.0 0.07

105 92.0 0.08

106 90.5 0.05

107 91.0 0.04

●​ Worked Solution:
1.​ Stress Conditions: We need to identify time steps t where both conditions
are met:
■​ Condition 1: V(t)>Vthreshold​=0.05
■​ Condition 2: ΔP(t)=P(t)−P(t−1)<−ΔPthreshold​=−1.5
2.​ Analyze Step-by-Step:
■​ t = 101: V(101)=0.03≯0.05. Condition 1 fails. (Need P(100) for Condition
2).
■​ t = 102: V(102)=0.04≯0.05. Condition 1 fails. ΔP(102)=97.0−98.5=−1.5.
Condition 2 is met (or borderline, depending on strict inequality).
■​ t = 103: V(103)=0.06>0.05. Condition 1 holds. ΔP(103)=94.5−97.0=−2.5.
Since −2.5<−1.5, Condition 2 holds. Stress identified at t=103.
■​ t = 104: V(104)=0.07>0.05. Condition 1 holds. ΔP(104)=95.0−94.5=+0.5.
Since +0.5≮−1.5, Condition 2 fails.
■​ t = 105: V(105)=0.08>0.05. Condition 1 holds. ΔP(105)=92.0−95.0=−3.0.
Since −3.0<−1.5, Condition 2 holds. Stress identified at t=105.
■​ t = 106: V(106)=0.05≯0.05. Condition 1 fails (borderline).
ΔP(106)=90.5−92.0=−1.5. Condition 2 is met (or borderline).
■​ t = 107: V(107)=0.04≯0.05. Condition 1 fails. ΔP(107)=91.0−90.5=+0.5.
Condition 2 fails.
○​ Result: Based on the specified criteria, periods of market stress occurred at
time steps t=103 and t=105.
○​ Interpretation: This type of analysis allows researchers to identify specific
regimes or events (like crashes or high-stress periods) within the complex
dynamics generated by the ABM. Studying the conditions leading into and out
 of these periods, or the agent behaviors during them, can provide insights
into the mechanisms driving emergent market phenomena like financial crises,
which may not be captured by simpler equilibrium models.

Exercise 30: Machine Learning Model Formulation (SVM for Prediction)

●​ Problem: A research study proposes using a Support Vector Machine (SVM) to
predict the direction of the next day's stock price movement (Up or Down). The
prediction is based on three features calculated from the previous day's data:
○​ X1​: Normalized return of the previous day.
○​ X2​: Normalized trading volume of the previous day.
○​ X3​: A measure of market sentiment (e.g., derived from news). Let the target
variable yi​be +1 if the price goes Up on day i, and −1 if it goes Down. Briefly
describe the objective of a linear SVM classifier in this context. Write down the
form of the decision function and the basic idea behind finding the optimal
hyperplane. (Do not solve the full optimization problem).
●​ Worked Solution:
1.​ Objective: The objective of a linear SVM classifier is to find the optimal
hyperplane (a line in 2D, a plane in 3D, or a hyperplane in higher dimensions)
that best separates the data points belonging to the two classes (Up=+1,
Down=-1) in the feature space defined by X1​,X2​,X3​.
2.​ Feature Vector: For each day i, we have a feature vector xi​=(X1,i​,X2,i​,X3,i​).
3.​ Decision Function: A linear SVM uses a decision function of the form:
f(x)=w⋅x+b where w=(w1​,w2​,w3​) is the weight vector (normal to the
hyperplane) and b is the bias term (related to the offset of the hyperplane
from the origin).
4.​ Prediction Rule: The predicted class for a new data point x is given by the
sign of the decision function:
■​ Predict Up (+1) if f(x)≥0.
■​ Predict Down (-1) if f(x)<0.
5.​ Optimal Hyperplane: The "optimal" hyperplane is the one that maximizes the
margin between the two classes. The margin is the distance between the
hyperplane and the nearest data points (called support vectors) from either
class.
6.​ Basic Idea of Optimization: Finding the optimal hyperplane involves solving
a constrained optimization problem. The goal is typically formulated as
minimizing the norm of the weight vector, ∣∣w∣∣2, subject to constraints that
ensure all data points are correctly classified with a margin of at least 1 (for
the linearly separable case): yi​(w⋅xi​+b)≥1for all training points i In practice,
data is often not perfectly separable, so a "soft margin" SVM is used, which
 allows for some misclassifications by introducing slack variables and a penalty
term in the objective function.
○​ Interpretation: By finding the hyperplane that maximizes the separation
between historical instances of "Up" days and "Down" days based on the
chosen features (X1​,X2​,X3​), the SVM aims to create a decision boundary that
generalizes well to predict the direction of future price movements. The
effectiveness depends heavily on whether the chosen features contain actual
predictive information about future price direction.

Closing Remarks
The exercises presented herein cover a diverse range of topics, from fundamental
physical principles to their sophisticated application in quantitative finance. They are
intended to foster a deeper understanding of the mathematical models,
computational techniques, and conceptual frameworks discussed in the associated
research context. By working through these problems, users can gain proficiency in
applying these methods and develop a critical perspective on their assumptions and
limitations when modeling complex systems, whether physical or financial.