Machine learning for quantitative trading.

KHH Analytics researches, builds and runs systematic strategies across global markets. We judge every strategy by what it returns per unit of risk taken and focus on keeping volatility low at all times.

Start a conversation See our methodology

  • Signal Research

    Every signal starts with an economic reason and a small set of features we can explain. Nothing reaches the book just because a search turned it up.

  • Systematic Execution

    The system generates the trades and keeps the book hedged against risks we are not paid to take. We do not step in to change what it decides.

  • Risk-First Construction

    Every allocation is sized against drawdown, correlation and exposure to changing market conditions before it is sized against expected return.

01 — The Thesis

Markets are efficient in aggregate and inefficient at the margins.

The strong form of the efficient market hypothesis has never survived contact with real data. What survives is something more useful and considerably harder to exploit. Markets are mostly efficient, most of the time, for most participants. Whatever is left over is faint, non-stationary and buried in noise.

The usual response is to throw scale at the problem: engineer hundreds of features, search for whatever predicts and keep the candidates that survive an out-of-sample test. That recipe defeats itself. Every rejected candidate reveals a little more about the holdout data, so after enough attempts the survivors have been selected on that data as surely as if they had been trained on it. Out-of-sample evidence is a finite resource and most quantitative research spends it long before a model ever reaches a live book.

We work in the opposite direction. We purposefully limit the number of features we use, both to avoid overfitting and to make sure our models do not break when market conditions change. At the same time we ask the right questions to find true inefficiencies hidden in the noise. We use alternative datasets to model relationships between assets and rank them based on simple and explainable features.

02 — Methodology

Every model starts with a reason to exist.

We would rather test one idea we can justify than screen a thousand we cannot.

  1. 01

    Data Engineering

    Price, fundamental and alternative datasets are ingested into a point-in-time research environment free of survivorship and look-ahead bias. Every model sees only what was genuinely knowable at the moment of decision.

  2. 02

    Signal Research

    Every signal begins as a hypothesis with an economic reason to exist, not as the output of a search. We keep the feature set small and the model simple enough to explain, then test it on periods it was never fitted to. Ideas that need complexity to work are dropped rather than rescued.

  3. 03

    Portfolio Construction

    Forecasts become positions through a process that models transaction costs, market impact and the correlation structure of the book. What gets optimized is the portfolio, never the individual trade.

  4. 04

    Monitoring & Decay Control

    Live models are continuously measured against their research-period distributions. The portfolio then adjusts risk automatically if strategy performance differs from what is expected. This way, decaying strategies are removed from the portfolio before they do real damage.

03 — Philosophy

Risk-adjusted performance is all that matters.

Raw return is trivially easy to manufacture: add leverage, concentrate exposure, sell volatility and wait. Each of these will outperform a broad index for a period of time, and each will eventually surrender that outperformance in a single adverse move.

We evaluate every strategy on what it earns per unit of risk taken: Sharpe and Sortino ratios, maximum drawdown, time to recovery, tail-loss behavior and the stability of those measures across independent out-of-sample periods and market regimes. A model that beats the market with twice the volatility has not beaten anything.

There is plain arithmetic behind this. A portfolio that falls 50 percent has to double just to get back to where it started, and the years spent climbing out are years it is not compounding. Keeping volatility low therefore reduces our drawdowns and, for the same average return, compounds to more over time.

04 — Contact

Discuss your mandate.

We work with allocators, family offices and asset managers seeking systematic, evidence-based exposure. Reach out to begin a conversation.

contact@khhanalytics.com