Want to understand how the best models in Numerai crush the competition every era? 🧵

Cue: Feature Exposure

Feature Exposure is a diagnostic metric measuring how much your model depends on specific features.

High exposure creates regime-dependent performance and should generally be reduced.

Feature Neutralization is the technique for reducing exposure by removing linear feature relationships from predictions.

It:
1) Uses pseudoinverse projection to residualize predictions
2) Preserves higher-order interactions while removing direct effects
3) Can be tuned via proportion parameter and feature selection

The Core Trade-off:
Lower exposure = more consistent performance but potentially lower correlation. The goal is finding your optimal balance.

Best Practice:
Selectively neutralize to the riskiest features with a tuned proportion parameter, rather than fully neutralizing to all features.

---
This is the TLDR of feature exposures.

Liked what you read?
Let's strike a deal.

> You RT
> I get 100 retweets
> I publish the full write-up below this (1/N)

Feature exposure is one of those concepts that separates "I have a backtest" from "I have a strategy."

It measures how concentrated your model's reliance is on specific features. Calculation is just Spearman correlation between predictions and each feature.

High exposure = you're betting those features continue to work. Low exposure = your features have no linear relationship to the target (might be noise).

Why this matters:

1. Markets are non-stationary. Features that crush it in one regime can be worthless in the next. If your max feature exposure is too high and is concentrated in too little features, you are betting that "these features are going to be good forever".

2. When regime shifts hit, high-exposure models don't "underperform" - they burn. The original Numerai forum post on this topic is literally someone describing getting destroyed in round 223 after overfitting on a limited feature set.

3. You're interested in Mean exposure AND Max exposure. Max exposures tell you how "reliant" you are on a single feature. Mean exposures, along with other aggregations, give you a sense of your distribution of "dependency". Heuristic: Max exposure of ~0.29 is considered reasonable.

The trap most people fall into:

They see "reduce feature exposure" and try to minimize it to zero. Now they have a robust model that predicts nothing.

The actual goal is finding the optimal trade-off between:
- Robustness (consistency across regimes)
- Predictive power (correlation with targets)

You want a model that doesn't blow up when markets change, but also actually makes money when they don't.

First Principles Of Multi-Manager Funds🧵

Why Do PMs/Traders Join Podshops?

1. (GOOD) They want access to capital/resources

Capital raising is a separate skill.

Some PMs can barely looking someone in the eye, let alone hold a conversation convincing someone to part with their hard-earned money. Sometimes, these PMs instead have some skill at generating alpha, and want to monetize the excess capacity of their alpha generation skills. So, they look for someone who can actually raise capital, and this makes a pair as old as time. These are the best relationships because both parties know they can't do without the other.

Separate from the above, sometimes, some strategies require a lot of resources to run them. These resource intensive strategies are typically because:
a. Data can be expensive [Many >100K subscriptions]
b. Storage can be expensive [TBs of data]
c. Compute can be expensive [GPU/High Ram]
d. Access can be expensive [Exchange seat, colo, etc]
e. Researchers can be expensive

They have strategies that work, but are conditional upon some resources that they individually cannot afford. So once again, the guy that can make money from money pairs up with the guy who can raise money to spend money to make money - pair made in heaven.

2. (BAD | GOOD) They want/need access to infrastructure

PMs don't emerge fully formed. Many of them are researchers (non risk-taking) looking for PM (risk-taking) seats. They will come with incomplete mental model of how the world works and are mostly derivative. Most of these are lemons, and are bad copies of their previous masters.

I'm sorry if this sounds harsh, but this is the truth. Researchers are MOSTLY low resolution JPEGs of their PMs, some of these are going to be some hilarious 8bit representations (e.g. Asia Index rebal researchers not knowing what is ToyoKezai).

How low resolution depends on:
a. How long they have spent with the PM
b. How much exposure they have with the PM
c. How much does the PM share with them
d. The culture of the team
e. Their propensity of "putting the pieces together"

Occasionally, you will find a researcher that has a unique worldview and is NOT a lemon. He has been able to synthesize the methodologies of his PM and via exploration/introspection, have something truly value-adding or novel to the process. Often times, the best marker for this is very very high ownership (hopefully single owner) of a complex but ultimately successful project; and being able to articulate and break down this project in high levels of detail in conversations.

Still, these researchers/first-time PMs will need help.

In the investing process, there are
-> DATA ONBOARDING
-> RAW DATA
-> CLEANED DATA
-> FEATURES
-> SIGNALS
-> PORTFOLIO
-> EXECUTION

In the tech stack there are:
- DATA ETL
- DATA STORAGE
- EXECUTION/OMS
- DATABASES
- MESSAGING BUSES
- CLOUD COMPUTE
- SIMULATORS

As a researcher, most of their time will be spent at the CLEANED DATA -> SIGNALS generation stage, with almost no exposure elsewhere.

It is not uncommon to hear first time PMs explicitly asking for help with :
DATA ONBOARDING -> CLEANED DATA
PORTFOLIO OPTIMIZERS
EXECUTION ALGORITHMS
PUT TOGETHER A TECH STACK

PMs generally really favor help with both ends of the investment process AND a modular tech stack. Generally, they want EXACTLY that and nothing more - because they want to obfuscate as much of their known worldview as possible: CLEANED DATA -> SIGNALS; and who can blame them?

It is, after all, their only advantage and bargaining chip.

3. (BAD) They want to be employees earning a stipend

Sometimes, PMs, like every other career, are just trying to make a wage to go home to their wife and play with their kids. They don't really care about anything else, but have realized that this industry is an EXCELLENT place to be paid to do absolutely nothing. Are they winners or losers? This truly depends on who you ask.

Some people call these guys CONSTANT GARDENERS. As a podshop, you generally want to avoid these guys like the plague. They are going to be pure cost centres, and are ONLY distractions to the firms.

Avoiding them is actually very easy if you are willing to accept a high false positive rate.

Avoid PMs that haven't been able to stay at a place for more than 2 years more than twice. It takes almost an entire year to build out a proper team/book, so that's the lifeline of most podshops. If the PM has been let go shortly after a year more than once, he likely hasn't got it; and is just farming gardens/wages.

The more name brands he has collected on his resume, the MORE likely he is a constant gardener. Let me put it this way in no ambiguous terms: Podshops will literally suck d*ck to keep a great PM. Only bad PMs get to leave, and a super-star PM getting poached is going to be known to EVERYONE. So the only surprises you get are going to be negative.

When in doubt, avoid >first-time PMs. The chances of him suddenly "developing edge" is close to 0. And for every washed up PM, you can pick from 5x researchers.

In fact, I am staunchly of the opinion that the only space in the industry in mature markets is exercising manager selection alpha on first-time PMs. (2/N) First Principles Of Multi-Manager Funds🧵

Why Do Podshops Want PMs/Traders?

In a single word, diversity:
I think you can think of the PM diversification axis as having (broadly) these dimensions that you want to focus on:

Data
x Style | Strategy
x Strategy Implementation
x Asset Class
x Region
x Universe

-----
You can unfold this into a big 2D map where you are essentially thinking along the lines of:

(Data x Style x Implementation) Along the Y-axis
(Asset Class x Region x Universe) Along the X-axis

This effectively gives you a diversification map of: WHAT IS BEING RAN (Y) x WHERE IS IT BEING RAN (X)

And every PM with capital deployed is a point on the map.

-----
Essentially, all Podshops want to:

1) Populate as much of this chart as possible. You want to (if resources were unconstrained) go as far as possible across the map (and it is a VERY big map). The reason why you want to do this is because, plainly, diversification increases the sharpe of the firm AND at the same time increases the capacity of the firm. Both are extremely desirable qualities for a money manager.

2) Watch out for any hotspots where you are concentrating capital on. Sometimes, if you're not careful, you can wind up with capital hotspots where you are concentrating too heavily on an area of the map. Then, you WANT to spread this capital more thinly to other areas where you have PMs, because you now end up having significant concentration risks. BUT, if you don't have enough PMs scattered across the map, you are forced to contend with your capital hotspots.

-----
Reality is a little more nuanced than this, because of 2 competing optimizations:

a. You want sufficient PMs in (Asset Class x Region x Universe) because you want to fully utilize netting benefits AND

b. Expertise/competitive advantage in a market takes time/luck for a firm to grow. Sometimes you can hire a PM that can effectively "unlock" an entire area of the map for the firm, but most of the time, if the firm does not have existing knowledge/expertise in an area of the map, hiring (junior) PMs there is going to challenging at best and throwing good money after bad at worst.

-----
We will eventually cover how to screen for PMs specifically that can unlock dark areas, but for now, the above brings us to the next point:

3) PMs are also hired to densify (Asset Class x Region x Universe). Trading is expensive, and the surest way to bring trading costs down is not to pay it at all. As far as I can tell, the only way to not pay for trading costs is to ensure that your orders never reach the market. Netting is essentially this - you have X number of teams, and what you do, is take all of their orders together, and pretend for a second that you are the exchange. You cross aggressive orders against passive orders, and attribute fills internally. Then, you take residual orders and only send THOSE to the exchange. In some firms, this can save up to 80% of trading costs.

4) Netting also has the additional benefit of ensuring that only your best stuff are sent out to market. Suppose you have 5 different teams, and 4 of them wants to buy apple and 1 of them wants to sell apple, netting allows you to perform a "dollar weighted voting" at the firm level for what the final positions should be, which decreases the variance of the firm significantly, since you only take on positions that are, by definition, "the best".

5) As a fun loop around, netting benefits are really ONLY observed when correlation among PMs are low. Which means that, you want to look for PMs that are in the same (Asset Class x Region x Universe) and yet have sufficiently different (Data x Style x Implementation) to introduce diversification in their positions.

A thread on correlation thresholds for signal submission...

1/8 Many teams have a signal pool where you can submit signals as a QR. Before you submit your signals, you often have to go through some threshold checks. Common thresholds are variations of sharpe, correlation, turnover.

2/8