Past Meetup

Handling data in machine learning strategies, Dr Michael Steele, Wharton School

This Meetup is past

40 people went

Price: $5.00 /per person

qplum office

185 Hudson st, suite 1620, plaza 5 · Jersey City

How to find us

Five minutes walk from Exchange Place PATH station. Plaza 5

Location image of event venue


This will be an interactive session. We will try to cover as many of the following topics as time permits:

1. What makes a Monte Carlo Model Acceptable?
a. A widely used model which does offer some insight and which people at the top of the game should be reluctant to accept.
b. Candidate Idea: A model is “acceptable” if the series generated by the model and the historical series of returns cannot be distinguished by a “fair rule”
i. What are fair rules?
ii. What are unfair rules?

2. Volatility Drag: How Does it Hurt, or Help?
a. 57 Varieties: Average Returns, Expected Returns, Realized Returns, Compounding Rates
b. Volatility is “worse” since returns are not independent
c. How can this guide money management choices

3. Things Change. Can one draw guidance from past paradigm shifts?
a. History of the Black-Scholes model (bad fit pre-80’s, then good fit, then bad fit post-87 --- not used to fit --- now)
b. Nixon and Gold
c. Japan Deflation
d. Negative interest rates
e. Oil Shock
f. One-quarter, one-eight, one cent
g. Planning for the “next shift”

4. Really Lucky or Really Good?
a. The curse of multiple comparisons
b. The Bonferoni Rule
i. Mathematically righteous, but
ii. So conservative as to crush scientific progress.
c. New ideas about false discovery rates---do they help.
d. Martingales and “Beating Warren Buffett”
i. Observed vs unobserved risks.
ii. A strategy that expects to beat the market by 5% for 20 years.
iii. Foster-Stine martingale benchmark
e. Incompletely Explored Mystery: Maximizing expected returns seems to get us into trouble less often than one might expect. Why?

5. What do you mean risk?
a. Even if we agree on “standard deviation” this is not as well-defined as one might think.
b. The logical tangle of “permanent loss” vs “temporary loss”. There is a mean-reversion pony buried in there someplace.
c. Drawdown has a visceral appeal but no common standards.
d. Enough risk to act?
i. What causes a client to change managers?
ii. Retail clients vs Institutional Clients

6. Is “everything” really just “regression”? Simple regressions are the “strawmen” of choice. They are surprisingly hard to beat, and it’s a big deal when you can beat them. Still, you have to set things up right.

Mike (J. Michael Steele) is C.F. Koo Professor of Statistics Emeritus of the Wharton School of the University of Pennsylvania.

He has long been involved with data-driven financial services, the analysis of financial time series, and the applications of machine learning to financial markets. Before Wharton he taught at Princeton University, Carnegie-Mellon, and Stanford University.