Scribd
Upload
357 views25 pages

Algorithmic Trading & Quantitative Strategies Gappy Lecture 2

Uploaded by

Anshuman Ghosh
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
357 views25 pages

Algorithmic Trading & Quantitative Strategies Gappy Lecture 2

Uploaded by

Anshuman Ghosh
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 25

Algorithmic Trading &

Quantitative Strategies
Lecture 2 (2/13/2024)
Giuseppe Paleologo (gardener)
Today's Remote Session
● I will upload the slides and the course notes tonight after class
● There will be a short multiple-answer quiz at the end of the
lecture.
● Please send answers in the form
○ 1-a
○ 2-c
○ …

To [email protected]. You will have 5 minutes


● I will post a homework Thursday
● Office hours+Q&A: if it is ok with you do them over zoom, I can
schedule two 1-hr sessions at 7PM on Mondays and Thursdays
Topics

● Recap from Last Class


○ GARCH
○ Kalman
● Introduction to Linear Models
○ Interpretation
○ Operations
○ Applications
Kalman
Kalman Filter

Define

⇒
Formulas

(Kalman Gain)

(Basic Formulas)

In steady state:
(Riccati Equation)
The Simplest Example
A More Complex Example
A lognormal model for returns:

Trick:

Observation:
Linear (Factor Models)
alpha systematic

idio rets

factor
loadings
factor
returns

The model has extra degrees of freedom


A feature, not a bug
Interpretations of B
Types of Factor Models

1. Fundamental: you know the loadings but not the


factor returns
2. Statistical: you don't know the loadings nor the
factor returns
3. Macroeconomic: you know the factor returns but not
the loadings
Important: Alpha Spanned and Alpha Orthogonal

Write

Choose

Alpha orthogonal is extremely


valuable. Very high SR
Transformations #1: Rotations
Replace loadings and returns with a "rotated" version:

You get an
equivalent model:
Examples:

1. Uncorrelated, unit vol factor returns


2. Orthogonal loadings
3. Can we center loadings and get an equivalent model?
Transformation #2: Projections
Sometimes, the model is just too big. Why?

1. Maybe it's good, but hard to interpret


2. Maybe it's not even that good (=commercial models) and it's just
better to have a simpler model?
From To
Transformation #3: Push-outs
● This transformation increases the dimensionality of the model
● Posit that the residuals are not uncorrelated and/or can be
predicted

● The essential step is "orthogonalization" of A with respect to B:


regress A on B and keep the residuals, so that B'A=0
● This is the same as "Multivariate regression as a sequence of
univariate regressions" (read Friedman-Hastie-Tibshirani)
● And the same as Cholesky
● Read the Appendix: Frisch-Waugh-Lovell is essential
Applications #1: Performance Attribution
Gives you an understanding of where the PnL comes from

"Factor exposures"
Application #2: Risk Management
Start with variance prediction and decomposition:

But then you can answer almost any question:

1. First order of concern: take first derivatives of everything


2. Strategic risk management: create scenarios where you
a. Stress parameters
b. Stress future returns
Application #3: Alpha Research
1. Linear models are powerful
2. They are not only linear. Nonlinearity is hidden inside the
factor loadings. What 99% of research think of as non-linear is
just linear models with arbitrary loadings
3. A better description is "shallow models". Shallow models are
great, except when dating, probably
4. Recent research on "benign overfitting" extends models to very
large-dimensional models
Application #4: Portfolio Management
1. Factor models fit like a glove with mean-variance optimization
2. Example: it's easier to address data errors with factor models
3. And it's easier to interpret and solve these models
Quiz 4. When pushing out a model, you build
a model using the
1. Can you model autoregressive a. Residual returns of the existing
model
processes order greater than 1 with
b. Total returns
Kalman Filters? c. Either works
a. Yes, they are just bigger
5. Performance attribution decomposes
b. No, because it's only for linear
a. The PnL of a strategy
processes
b. The return of a strategy
c. No
c. Both
2. What types of alphas is related to
6. What can't you use a factor model
factor returns?
for?
a. Alpha orthogonal
a. Alpha research
b. Alpha Spanned
b. Market impact
c. Both
c. Risk management
3. In a factor rotation the matrix C
7. What model is used most often in
must be
practice?
a. Unitary
a. Macroeconomic
b. Nonsingular
b. Fundamental
c. Symmetric positive definite
c. Statistical

You might also like