Scribd
Upload
318 views22 pages

Algorithmic Trading & Quantitative Strategies Gappy Lecture 5

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
318 views22 pages

Algorithmic Trading & Quantitative Strategies Gappy Lecture 5

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/ 22

Algorithmic Trading &

Quantitative Strategies
Lecture 5 (4/2/2024)
Giuseppe Paleologo (gardener)
Today's Session
● I will send a project description on Thursday
● 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


Recap on Last Quiz
1. A drawback of statistical models is 4. Why can't you use Marchenko-Pastur
a. Low predictive power for cleaning correlation matrices?
b. Computational burden a. Unit variance assumption does not
c. Limited history hold
d. Low information use b. Zero correlation assumption does
2. The standard PCA is based on an not hold
underlying factor model c. Sample is too small
a. True 5. PPCA recommends to
b. False a. Shrink the factor volatilities and
c. Uncertain correlations
3. If two eigenvalues of a matrix are b. Shrink the factor correlations
identical, its eigenvectors have c. Shrink the factor volatilities
a. Indefinite norm 6. PPCA recommends to
b. Indefinite direction a. Shrink the idio volatilities
c. The jargon name "Dick" and "Jane" in b. Shrink the big idio volatilities,
the academic literature increase the small ones
d. Very important c. Leave the damn things unchanged
Topics
● Statistical Models
○ Practical considerations: z-scoring and reducing turnover
● Fundamental Factor Models
● Next Time:
○ Recap on Fundamental Models
○ Advanced Portfolio Construction
Why Fundamental Models?

● Interpretable
● Extensible
● Good performance
● Connects to academic research and to alpha research
Primitives
● Returns
○ Which returns? Closing auctions. Based on equal weighted/EWMA/volume
weighted averages
● Characteristics
○ "Features", "Terms", "Signals"
○ Continuous variables
○ Discrete/categorical
Process
● Data ingestion and quality assurance
○ Missingness
○ Turnover
● Winsorization
● Estimation universe selection
● Loadings generation
● Cross-sectional regression
● Time-series estimation
Statistical Models
Reminder: positive features of the spectrum of a matrix are:

● The spectrum is high and separated from the bulk. The spiked
assumptions hold better
● Spike eigenvalues are separated. The eigenvectors are
identifiable
Whitening with idio vol
proxy helps
Additional
Benefits of Idio
Whitening:
Returns
Additional Benefits of Idio Whitening: Returns
Reducing Turnover
Factors are turn over a lot. They can also change sign. They can be
stabilized by rotating them from one period to the next.

There is a clever solution for this problem.


Read the notes for details
Putting it all Together
Fundamental Models
Takeaways for Fundamental Models

● At their core, very simple: at their core, alpha and risk modeling
are very simple
● There are many details
○ Data
○ Shrinkage
○ Heteroskedastic and correlation correction
○ Regime Change
Steps
Steps:

1. Data ingestion: integrity, temporal continuity, missingness,


outliers
2. Estimation universe: data quality, continuity, liquidity,
relevance
3. Winsorization
4. Loadings Generation
5. Cross-Sectional Regression
6. Covariance Generation
Cross-Sectional Regression

Assumptions:
1. Loadings are full rank
2. Homoskedastic idio returns are homoskedastic (jkg)
3. Factor returns and idio returns are independent
4. At least fourth moments are finite
Basics of X-sec regression
Gaussian Likelihood

Note the similarity with PCA


Note that the estimation is exactly the same as factor mimicking portfolios
And this is why you should use squared loss
Shrinkage
● Factor volatilities are biased upwards because they are
regression coefficients with errors, so they should be shrinked:

● Anderson's asymptotics for T→∞ doesn't apply. Recommended:


simple Ledoit-Wolf. Let's not overkill.
Separation Volatility-Correlation
Correlations change at a slower time-scale than volatilities
When people say, "in a crisis all correlations go to 1"... they should
think twice [discuss!]
Dynamic Factor Volatility
Add a "modulating term".

The intuition is that


cross-sectional factor returns
are pretty large, and
responsive.

If you get them wrong, adjust.


Autocorrelation Correction
You need this. Example: Asynchronous Returns.

Scholes correction:

Newey-West correction:
Winsorization
● Ideally, choose the right estimation/investment universe and
never winsorize
● In practice, the investment universe selection is imperfect
○ Bad data sources
○ Illiquid assets
○ Bankruptcies relistings
● So use a simple measure

You might also like