Where an idea comes from

Every hypothesis starts with an observation in the data, not with a view on where the market is heading.

We work from end-of-day price histories, option chains with their implied volatilities, realised volatility measured over several horizons, and volatility indices. From these we derive the quantities that describe an options position: the spread between implied and realised volatility, the shape of the term structure, the distribution of returns over the intended holding period.

An observation only becomes a usable hypothesis once it is stated precisely enough to fail. That means naming the instrument, the entry condition, the holding period and the exit rule before any testing begins, together with the result that would make us discard the idea.

A hypothesis that cannot be refuted by any conceivable outcome is not a hypothesis. It is an opinion.

• Data basisPrice and option histories spanning full market cycles, including periods of stress.
• Stated in advanceEntry, holding period, exit and the criterion for rejection are fixed before the first test.
• Discarded earlyMost ideas do not survive this stage, which is the point of it.

Testing against more than one history

Realised history is a single sample. A strategy that only works on that one path tells us very little.

We first test over long historical windows that include several distinct volatility regimes: calm trending markets, sharp shocks and the extended recovery phases that follow them. A strategy has to remain acceptable in each of them, not merely profitable on average.

We then run Monte Carlo simulations. Instead of the single price path that actually occurred, we generate a large number of alternative paths with comparable statistical properties and evaluate the strategy on all of them. This produces a distribution of outcomes rather than one number, which is what allows us to reason about the bad tail instead of the average case.

Sensitivity analysis closes the stage. We vary strike distance, holding period, entry filters and cost assumptions and observe how the result responds. A strategy whose performance collapses under a small parameter change is fitted to the past rather than describing a persistent effect.

What the models do, and what they do not

We use machine learning where it is genuinely better than a fixed rule: recognising patterns across many interacting variables.

In broad terms we work with supervised models on engineered market features, ensemble methods for estimating probabilities, and classification of volatility regimes. The purpose is not to predict prices. It is to estimate conditional quantities: how likely a given position is to reach its target under current conditions, and how wide the plausible range of outcomes is.

Three uses in particular. Risk assessment, where a model weighs current market conditions against comparable historical states. Strategy refinement, where parameter choices are evaluated across conditions rather than tuned to a single average. And forecasting expected trade outcomes, stated as a distribution with its uncertainty, never as a single figure.

The division of labour is deliberate. Models deliver analysis, ranges and confidence levels. Every decision to open, adjust or close a position is taken by a person, against the documented rules of the strategy. No model has order-execution authority, and a model output that contradicts the risk framework does not override it.

A model that cannot explain why it is confident is treated as one input among several, not as an answer.

Small first, and always with a defined loss

A validated strategy goes live in the smallest size that still produces meaningful information.

Position sizing follows from the risk budget, not from conviction. Each strategy is allocated a share of capital and a maximum loss per position, and both are set before the first trade. Size increases only after live behaviour has matched the modelled behaviour over a meaningful number of trades.

Exit rules are written down in advance: profit target, maximum loss, time-based exit and the conditions under which a position is adjusted rather than closed. Because they exist on paper before the trade, they do not have to be invented while the market is moving.

• SizingDerived from the risk budget per strategy and a fixed maximum loss per position.
• LimitsMargin and stress scenarios are checked before entry, not after.
• ExitsTarget, stop, time exit and adjustment conditions documented in advance.

Watching for the moment the model stops fitting

Live monitoring exists to detect divergence between modelled and actual behaviour early enough to act on it.

Exposure, decay, volatility sensitivity and margin usage are reviewed daily, alongside the volatility regime the strategy assumes. Realised results are compared against the distribution the simulation predicted. Individual losses inside that distribution are expected and change nothing; a systematic deviation from it does.

When behaviour drifts outside the modelled range, the response is graded: reduce size, adjust parameters, or stop the strategy and return it to research. Stopping is a normal outcome, not a failure of process.

We are explicit about how we measure and decide, including where results fall short of expectation. Any performance figure without its method, its period and its risk taken is not information.