A mechanism before a test
No hypothesis enters the backtest harness without a written economic reason for why the effect should exist. This is the cheapest filter we have. A signal that only has a p-value has nothing.
Page 01/05
Approach
How we think, not what we hold.
What follows is the methodology, at the level a serious reader needs to judge whether the process is sound. It is deliberately not the parameter set. The signal logic is the part worth keeping private; the discipline around it is the part worth being judged on.
- Universe
- Midcap 150 + Smallcap 250
- Horizon
- Weeks to months
- Holdout
- One window, spent once
- Overrides
- None
- Mechanism first
- Point-in-time
- Spent-once holdout
- Failed searches count
- Costs inside the signal
- Constraints outrank conviction
01The stack
Where the edge is supposed to come from.
Four decisions, taken in order. Each one is a place where the obvious choice is worse than the careful one, and where the reason is a property of this market rather than a preference.
01
Selection
Cross-sectional, not directional
The desk ranks names against each other rather than forecasting the market. Momentum and risk-adjusted momentum do most of the work, with quality and liquidity screens acting as gates rather than as additional score components. Ranking on return divided by trailing volatility has been consistently better than ranking on raw return, which is the sort of small, dull result that tends to hold up out of sample.
02
Construction
Hierarchical risk parity over mean-variance
Covariance matrices estimated from a few hundred noisy mid-cap return series are not stable enough to invert, and mean-variance optimisers respond to that instability by concentrating into whatever the estimation error happened to favour. Hierarchical risk parity clusters the universe by correlation structure and allocates down the tree, which produces portfolios that survive the estimate being somewhat wrong.
03
Rotation
Sector tilts, sized by volatility
India's sectoral indices show one-month continuation rather than the reversal effect imported from US single-stock research. The rotation sleeve leans into that continuation, with sizing driven by realised volatility. We have tested and rejected regime-timing overlays on this sleeve three separate times, so the tilt is structural rather than tactical.
04
Risk
Vetoes, not penalties
Risk management here is a set of binary gates, not a term added to an objective function. A name that fails the liquidity screen is not down-weighted, it is removed. A price series with an unexplained single-day move is quarantined until the corporate action calendar has been checked. Soft penalties get traded away by a confident enough signal; vetoes do not.
02Research discipline
Six rules, and the fourth is the expensive one.
These are the rules that decide what gets deployed, and they are the reason most of what we test never does.
Slide 01/06
03Out of scope
Things we deliberately do not do.
A strategy is defined as much by what it refuses as by what it does. These are permanent exclusions, not a roadmap.
Intraday and high-frequency
We have neither the latency budget nor the microstructure infrastructure to compete there, and pretending otherwise would just be paying spread for the privilege.
Discretionary overrides
If a rule can be overridden when it feels wrong, the rule is not doing anything. Changes go through research, not through the order screen.
Language models in the decision path
They are useful for reading filings and triaging hypotheses. They are not permitted anywhere in signal, scoring, sizing, or execution, because those paths have to be reproducible from state.
External capital
The desk trades its own book. Taking outside money would change the regulatory position and the incentives, and neither change would improve the research.
Next
The infrastructure is the other half of the story.
None of the above works without a data layer that refuses to hand a signal something it could not have known at the time.