The range of work the firm takes on.

Any market where data can be obtained, research can be done and orders can be routed. Venue and jurisdiction are engineering constraints (data availability, market structure, session times, clearing and connectivity), not a limit on what work is accepted.

The constraints are concrete rather than abstract. Whether a venue publishes full-depth data or only top of book decides what microstructure work is possible at all. Whether its history is available point-in-time decides whether a backtest can be trusted. Tick size regimes, auction mechanics, closing procedures and self-match prevention differ enough between venues that a strategy validated on one and deployed on another is a different strategy. Session boundaries and holiday calendars decide what a trading day even means for a book spanning several regions, and getting that wrong is one of the more common sources of silent backtest error.

In practice a mandate is scoped to the venues a client already trades. The relevant question at the outset is whether the data and the execution path exist, not where they are.

Venue names describe the market structure research and systems are built against. They do not indicate membership, direct connectivity or affiliation.

Asset classes differ less in their pricing theory than in the practical structure around them, and it is the structure that decides whether a piece of research survives contact with production. Equities carry corporate actions, borrow availability and a fragmented venue landscape. Futures carry roll conventions, expiry mechanics and margin regimes that change the shape of a position without any decision being taken. Fixed income and credit are quote-driven and thinly traded, so the same statistical technique that works on a liquid equity index becomes a study of stale prices. FX has no consolidated tape, which means the definition of the price is itself a modelling choice.

Digital assets combine continuous sessions, fragmented venues, funding mechanics on perpetual contracts and a counterparty risk profile with no real analogue in traditional markets. Treating them as equities that happen to trade at the weekend is the standard way to get them wrong.

Non-linear instruments are where the argument in Method stops being decorative. A payoff whose value depends on the path taken, or on more than one underlying, has second-order structure that a first-order risk view does not capture. The diagonal of the risk matrix is reported by every system; the cross terms are where a book that looks flat turns out not to be.

Path dependence adds a second problem, which is that the risk is a function of history rather than of the current state alone. A barrier that has been approached and not breached is a different position from one that has never been near, even with identical spot, vol and time to expiry. Systems that carry only current state cannot represent that difference, and the error is silent.

Frequency is not a dial on a single strategy. It changes which part of the problem dominates. At long horizons the binding constraint is statistical: few independent observations, so almost everything rests on whether the evaluation was honest. At short horizons the constraint is mechanical: signal quality matters less than whether the order arrived, filled, and cost what the backtest assumed.

Systematic & relative value

Signal research, portfolio construction and risk models for holding periods measured in days to months. The scarce resource here is independent observations: twenty years of daily data on a slow signal is a few dozen effective bets, not five thousand. That is why the evaluation discipline matters more than the model, and why block bootstrap and walk-forward testing are used rather than treating each day as fresh evidence.

Volatility & derivatives

Surface construction and fitting, pricing and calibration, second-order risk on books with more than one underlying, and the hedging logic that follows. A calibration that fits today's surface tightly and moves erratically day to day is usually overparameterised rather than responsive, and the curvature of the objective is what distinguishes the two.

Market structure

Microstructure research, execution quality measurement and latency-sensitive order routing. Order book reconstruction from direct feeds such as CME MDP 3.0, Nasdaq ITCH and Cboe PITCH; queue position and fill-rate analysis; scheduling against VWAP, participation and implementation shortfall benchmarks. This is execution engineering rather than latency arbitrage: the firm builds routing and measurement; it does not operate a colocated trading business.

Cross-venue and cross-asset

Relationships that only exist between instruments: cash against futures basis, listed against OTC, the same risk expressed on venues with different microstructure and clearing. These are the strategies where the off-diagonal terms are the entire position rather than a correction to it, and where the operational failure modes (a leg that does not fill, a venue that halts, a margin call on one side) matter as much as the research.

Method selection is itself a second-order question. The right technique is the one whose assumptions match the structure of the data, and financial data has three properties that break most defaults: it is low signal-to-noise, it is non-stationary, and its observations are dependent. Any method chosen without accounting for those three will report a number, and the number will be wrong in a direction that flatters it.

Machine learning

Applied where the structure of the problem justifies it, which on financial data is less often than the literature suggests. The dominant failure is not model choice, it is evaluation. Random k-fold cross-validation on labels built from overlapping horizons leaks the answer across folds and produces a score that cannot be realised; purged and embargoed splits are the minimum, and even those overstate what a live system will see.

Feature and target construction carry more of the result than model class. How a label is defined, whether by fixed horizon, by a triple-barrier rule, or through meta-labelling a separate primary signal, changes what is being predicted far more than substituting gradient boosting for a neural network. Most reported improvements from a new architecture are improvements in the labelling that came with it.

Where it earns its place: high-dimensional problems with real and numerous interactions. Order-flow and microstructure features, alternative-data ingestion, execution scheduling, and fast approximation of expensive non-linear risk calculations. Where a regularised linear model and an honest evaluation scheme answer the question, that is what gets used, because a model that cannot be interrogated is difficult to operate and harder to defend on the day it stops working.

The curvature argument applies here directly and unkindly. A flat likelihood surface means the parameters were never identified; a model with many more parameters has many more directions in which to be flat, and almost none of them are inspected. Regularisation is a statement about which of those directions are believed to be uninformative, and that statement deserves to be made explicitly rather than inherited from a default.

Prediction markets

Treated as two separate things, and conflating them is a common error. As a research object, a prediction market is an unusually clean laboratory for questions about belief aggregation: how closely traded probabilities track realised frequencies, where the favourite-longshot bias appears and how large it is, how quickly a price incorporates news relative to other venues trading the same underlying uncertainty, and what liquidity does to all of the above.

As a signal source, they supply probabilities for events that are otherwise difficult to price: policy decisions, macro releases, elections, and discrete corporate outcomes. The interesting quantity is rarely the level itself but the basis, the difference between a prediction market's implied probability and the same event's probability extracted from an option surface or a rates curve. Two markets disagreeing about the same event is a more tractable research question than either one in isolation.

The caveats are structural and get stated up front: books are thin, capital is constrained, resolution criteria carry their own risk, and the regulatory position has moved repeatedly. Any research that depends on these venues is scoped with those limits explicit rather than discovered later.