EXTRACTING SIGNAL FROM NOISE
SYSTEMATIC FX TRADING · QUANTITATIVE RESEARCH
We build machine learning systems that detect non-random structure in financial market data, primarily currency markets. Proprietary pattern recognition combined with real-time statistical inference, capturing faint regularities across high-frequency price streams and converting fractional edges into systematic value.
PHILOSOPHY
We start with data, not theory. No preconceived notions about how currency markets "should" behave. Patterns must be validated repeatedly across large samples with rigorous out-of-sample testing. If something cannot be proven empirically, it is not pursued.
The scientific method applies without exception. Hypotheses are tested, not assumed. Results must be reproducible across different market regimes. Overfitting is the enemy; we optimize for generalization, not historical fit.
Small edges, consistently applied, compound. We are not looking for obvious signals; those are already arbitraged away. We are looking for micro-inefficiencies in FX markets that survive statistical scrutiny and can be exploited systematically at scale.
Thousands of precision trading decisions daily, each sized according to information-theoretic principles. The goal is not to be right every time, but to have positive expected value over a large sample with controlled drawdown.
RESEARCH
The Unreasonable Effectiveness of Humility in Quantitative Trading
On building a system that survives by admitting what it doesn't know.
THE PROBLEM
Currency markets generate massive tick-level data streams that appear random. Standard quantitative methods assume stationarity, Gaussian distributions, and independence. FX markets violate all of these assumptions, constantly.
Returns have fat tails. Cross-currency correlations shift during stress events. Regime changes occur without warning: central bank interventions, geopolitical shocks, liquidity crises. Methods that work in textbooks fail catastrophically when deployed. We build inference frameworks designed for these realities.
Black-box trading models that work without explanation are fragile. Prediction without understanding breaks under market stress. When conditions change (and in FX they always change), models with no interpretability offer no guidance on what went wrong or how to adapt.
We build models whose assumptions are explicit, whose failure modes are characterized, and whose behavior can be reasoned about analytically. Transparency is not optional; it is essential for surviving tail events.
APPROACH
DATA FIRST
We work with tick-level FX data at scale. Quality, depth, and latency determine what edges can be captured. Multi-venue price feeds, order book snapshots, execution analytics. Infrastructure is not an afterthought; it is foundational.
NO PRECONCEPTIONS
Patterns emerge from systematic analysis, not macro narratives. When data contradicts intuition about interest rate differentials or risk sentiment, the data wins. When a model contradicts price action, the model is wrong.
RIGOROUS VALIDATION
Every signal must survive out-of-sample testing, multiple testing corrections, and stress testing across different market regimes: risk-on, risk-off, high vol, low vol. The null hypothesis is always that the signal is noise.
SYSTEMATIC EXECUTION
Once validated, strategies execute without human override. Discretionary intervention during drawdowns does not improve properly validated systems. It degrades them. Discipline is encoded in algorithms, not willpower.
METHODS IN DEPTH
Advanced architectures for sequence modeling and pattern recognition in non-stationary FX markets. Our systems learn continuously and adapt to shifting market microstructure without catastrophic forgetting. This is critical when currency dynamics evolve through different liquidity regimes.
We focus on architectures that balance expressiveness with interpretability. Attention mechanisms reveal which price patterns and cross-currency relationships the model considers important. Ensemble methods quantify forecast uncertainty for position sizing.
Spectral analysis of FX price series reveals cyclical structure invisible in raw tick data. Wavelet methods decompose currency movements across multiple timeframes simultaneously, from intraday microstructure to weekly positioning flows.
Information-theoretic measures quantify lead-lag relationships between currency pairs without assuming linearity. Transfer entropy identifies which pairs lead during different market regimes. Optimal filtering extracts tradeable signals while characterizing estimation uncertainty.
Bayesian methods for estimating market parameters under uncertainty. Prior beliefs about volatility, correlation, and mean-reversion are updated with each tick. Posterior distributions quantify what we know about current market state and what remains uncertain.
Sequential Monte Carlo methods enable real-time estimation of latent market states: hidden liquidity, institutional positioning, regime probabilities. Particle filters track these as new price observations arrive, producing calibrated forecasts for execution.
Time-series models for FX processes with long memory, regime changes, and heavy tails. Standard GARCH models assume too much. Currency returns exhibit volatility clustering, structural breaks around central bank events, and flash crashes that violate Gaussian assumptions.
State-space frameworks handle time-varying parameters: correlations that shift during risk-off episodes, volatility that spikes during NFP releases. Hidden Markov models capture regime-switching dynamics between trending and mean-reverting states.
Entropy-based feature selection identifies which market variables carry predictive information for currency moves while discarding noise. Mutual information quantifies nonlinear dependencies between pairs that correlation matrices miss entirely.
Information-theoretic principles guide position sizing under uncertainty. Capital allocation follows from Kelly criterion and its extensions, optimized for long-term growth rate rather than short-term P&L. This distinction is subtle but critical for surviving drawdowns.
WHAT SETS US APART
Unlike discretionary FX traders or static quant models, our systems continuously learn and adapt to changing market microstructure. Dynamic optimization means the strategy improves over time, incorporating new market data while respecting what has been learned.
Regime-independent performance is the goal. Strategies that only work in trending or range-bound markets are brittle. We design for robustness across risk-on, risk-off, high-vol, and low-vol environments, accepting lower peak returns for more consistent risk-adjusted results.
Returns that are repeatable, scalable within liquidity constraints, and uncorrelated to conventional FX strategies. We are not competing on the same carry trades or momentum signals as everyone else. The edges we seek are structurally different: micro-structural, statistical, regime-dependent.
Capacity constraints are respected, not ignored. Some currency pair inefficiencies only exist at certain trade sizes. We size positions to match the available liquidity, not to chase AUM growth at the expense of alpha decay.
24/5 EXECUTION
Systems that trade continuously across all major FX sessions (Tokyo, London, New York), adapting to each session's distinct microstructure.
RISK-FIRST SIZING
Position sizing derived from information-theoretic principles. Kelly criterion optimization for long-term capital growth, not short-term P&L.
TAIL RISK AWARE
Models designed for fat-tailed FX returns. Flash crash scenarios, central bank interventions, and liquidity gaps are stress-tested, not ignored.