Approach
Process over prediction.
The investment process
Research, construct, control, review.
01
Research
Falsifiable ideas, tested out of sample.
- Every hypothesis states what would prove it wrong
- Train, validate and test on separate periods with purge gaps
- Ten seeds for any model with randomness
- Transaction costs and turnover included from day one
02
Construct
Allocate risk, not dollars.
- Balance risk across growth and inflation regimes
- Cluster assets by correlation; budget risk down the hierarchy
- Prefer robust estimators to fragile optimizers
- Liquidity sized to the worst week, not the average one
03
Control
Survive first, then compound.
- Volatility targets and exposure limits
- Fractional sizing: Kelly is a ceiling, never a target
- Pre-committed rebalancing and drawdown reviews
- Stress tests on historical and hypothetical shocks
04
Review
Write it down, especially the mistakes.
- Post-mortems on every closed decision
- Regular letters explaining positioning and risk
- Public research, including negative results
- Model changes logged and re-validated
Our research
Hierarchies, made learnable.
Hierarchical risk parity groups assets by how they move together and splits risk down that tree, avoiding the fragile matrix inversion behind classic mean-variance optimization. Our DHRP research makes the hierarchy differentiable, so it can be trained end to end, then tests it across eight asset universes with ten seeds, publishing where it works and where it does not.
Evidence standard
Six questions every backtest must answer.
Most backtests are flattering by construction. These tests are how we try to be less flattering to ourselves.
How to read a backtest like a skeptic →- Probabilistic Sharpe ratio
- Is the Sharpe ratio likely positive, given skew, fat tails and a short sample?
- Deflated Sharpe ratio
- Does it survive after accounting for how many strategies were tried?
- Probability of backtest overfitting
- Does the in-sample winner keep winning out of sample?
- Paired tests vs. baselines
- Is it better than simple alternatives, not just better than zero?
- Multi-seed error bars
- Would a different random seed have told a different story?
- Cost sensitivity
- At what trading cost does the edge disappear?
See the approach applied to your portfolio.