Transcript Rage Against the Machine (Learning)

Rage Against the Machine
(Learning)
R/Finance
20 May 2016
Rishi K Narang, Founding Principal, T2AM
What the hell are we talking about?
• What the hell is machine learning?
• How the hell does it relate to investing?
• Why the hell am I mad at it?
2
What the hell is machine learning?
•
•
•
Method for automating design of models by algorithmically studying data
Traditionally, model design is a human activity (e.g., first and second steps
of the Scientific Method)
Related (read: conflated) terms:
Data mining – attempts to discover previously unknown properties in data
• Artificial intelligence – sort of the parent field of ML. seeks to replicate (general)
intelligence within a computer. learning is one (very crucial) kind of intelligence
• Data science – umbrella covering all of these terms
•
•
Consider “data driven investing” instead of ML
3
No, seriously, what the hell is it?
• Supervised learning: non-parametric (model-free) input-output functions
• classification (e.g., Trees, SVM)
• regression (e.g., Gaussian processes)
• Unsupervised learning: non-parametric data representation
• clustering (e.g., k-means)
• dimensionality reduction (e.g., ISOMAP)
• density estimation (e.g., kernel density)
• Reinforcement learning:
• learning + dynamic control: learn to behave in an environment to maximize cumulative
reward
credit: Balasz Kegl
4
Ok, let’s try a different tack: What the hell are we
talking about when we talk about investing?
5
So what the hell do people usually do for
Alpha Models?
Return Category
What
Alpha
Price
Input Type
Phenomenon
Specification
How
Trend
Reversion
Forecast
Target
Fundamental
Technical
Sentiment
Model
Specification
Yield
Growth
Conditioning
Variables
Quality
Run Frequency
Time Horizon
High Frequency
Bet Structure
Directional
Instruments
Liquid
Long Term
Relative
Illiquid
Implementation
6
How the hell do you use machine learning to
forecast returns?
• What defines the current market condition?
• By what technique do you identify conditions and expected outcomes?
• What data should you (let the machine) study?
7
What the hell is the problem, exactly?
1. It’s really hard...
• very difficult to separate signal from noise, even with strong priors
• very difficult to prove your algorithm is doing what you meant it to do
...so most people attempting to utilize these approaches are simply not qualified
2. It’s a buzzword...
• my guess is that there are now ~100-200 quant funds claiming to utilize ML techniques, versus maybe 10 three years
ago
• investors are also very excited
...so much of what is being paraded about as “ML” is in practice just linear regression
• poseurs are annoying
3. Almost no one utilizing ML is successful
• especially in the alpha model itself (as opposed to the meta-alpha / signal combination phase) is successful
...so all the fuss is for no particularly good reason
HOWEVER, done well, ML has great promise as a way to discover subtler, less intuitive alphas
8