Build your own risk model
This site will take you through the process of building your own investment risk model from real, up-to-date data. Your own unique portfolio has been generated — a set of weights across selected UK equities, just like a real fund manager would hold, and you will see it as soon as you sign in.
Your task is to build various flavours of risk model that try to predict useful metrics like tracking error and volatility, in increasing order of complexity. You will download the data, use Excel to implement the techniques, and submit your estimates as you go — and this site will check that you are on the right track.
Start with Module 0, which explains what a risk model is for. There is no rush, and nothing here expects you to already know the answer.
Sign in to get your own portfolio
Each joiner is assigned a different set of holdings, so the self-checks grade your working rather than someone else's. No password — request a link and Charlie will forward it to you.
Sign inThe modules
-
0
What you are about to build
What a risk model is for, why we use Excel, and how to use this site.
15 min
-
1
The data, and what is wrong with it
Ragged histories, bad prints, survivorship bias, and the benchmark that is lying to you.
60 min
-
2
Returns, and the square root of time
Three frequencies, log versus simple, and why annualising daily vol does not quite match annualising monthly vol.
60 min
-
3
The covariance matrix
Build the whole matrix in one formula, then squeeze it with a weight vector to get portfolio volatility.
90 min
-
4
Tracking error and where the risk comes from
Active weights, and decomposing total risk into per-holding contributions that actually add up.
60 min
-
5
Time windows and exponential weighting
Your volatility estimate depends as much on the window you chose as on the market. Watch it move through 2008 and 2022.
75 min
-
6
Value at Risk
Parametric VaR is one formula. Historical VaR makes fewer assumptions. Then test whether either one worked.
75 min
-
7
The single-factor model
Regress every stock on the market, and rebuild the covariance matrix from 163 numbers instead of 3,321.
90 min
-
8
The cross-sectional factor model
Regress returns on stock characteristics each week to estimate factor returns.
2-3 hours
-
9
The statistical risk model
Extract factors directly from the covariance matrix using principal components. The first one turns out to be the market.
2-3 hours
-
10
The bake-off
Five models, one portfolio. Which would you have trusted in March 2020?
half a day
-
11
Skew, kurtosis, and a VaR that believes them
Measure how far returns are from normal, watch the measurement wobble when you change the sampling day, then feed it into a VaR.
75 min
-
12
Drawing a line into the tail
Plot the worst days on a log-log chart, fit a straight line, and extend it past anything in the data. Then ask how far to trust it.
75 min
-
13
Fat tails, crashing together, and Monte Carlo
Swap the normal for a Student-t, find out what correlation misses about stocks falling together, and simulate a portfolio that knows about both.
2-3 hours
Get the data
These links are here for reference. Each module links the download you need for it, so you should never have to guess which file to use.
Before you start
Work in Excel. Use whatever route you like — MMULT,
SUMPRODUCT, or a pile of tabs built on nothing but
* and +. The checker only validates the number.
And do ask the team how they would do it. You will get answers of varying degrees of over-engineering, so choose your fighter wisely.