We treat research as a self-improving system, not a sequence of one-off bets. Hypotheses, backtests, stress tests and live monitoring all live on the same substrate — so a good idea travels from paper to production in days, not quarters.
§ 01 — Strategy library
Each pillar inherits the same data backbone, the same backtester, the same risk system. We can mix, hedge and capacity-budget across them because they all speak the same language.
We rank thousands of correlated instruments, build long/short books to dampen margin pressure, and execute across multiple venues — capturing the mean-reverting noise between assets that should move together.
Relative-strength signals long the top decile and short the bottom. Short-horizon mean reversion captures the snap-back of recent over-extensions. Sized by realised volatility regime.
Momentum, value, carry, low-vol, quality — a factor library combined under a risk-aware optimiser that rebalances daily across regions and asset classes. Capacity-aware sizing keeps the book honest.
Kernel-bypass networking, custom exchange adapters and co-located gateways harvest fleeting price differentials across exchanges, dark pools and pairs of correlated derivatives.
Gradient-boosted ensembles, transformer-style sequence models and probabilistic graphical models trained on order-flow imbalance, options skew, news embeddings and on-chain signals.
Roughly 120 hypotheses enter the pipeline each week. Most fail — that is the point. The ones that survive backtests, stress tests and paper trading inform the next iteration.
§ 02 — Research lifecycle
No idea reaches production by intuition alone. The lifecycle below codifies the bar: every strategy on our book has cleared each gate, and continues to clear it on a rolling basis.
Read, conjecture, formalise. Papers in, hypotheses out — a dedicated harness summarises and structures the literature daily.
Vector-replayed against a bit-exact replica of historical microstructure. No look-ahead, no survivor bias, no rounding cheats.
Crisis windows — 2008, 2015, 2018 vol-spike, 2020 COVID, 2022 rates — and synthetic shocks. We size for the worst plausible week, not the median month.
Live data, real venues, zero risk. We compare to backtest in real-time and only graduate if the slippage profile matches expectations.
Promoted under capacity-aware sizing, monitored by a per-strategy dashboard, with kill-switches owned by independent risk.
When a strategy decays past its expected half-life, we retire it before it leaks. Discipline beats nostalgia.
§ 03 — Data substrate
We curate a single, deduplicated, bit-exact archive of market and alternative data — from tick-level order books to filings, news, options-implied surfaces and on-chain transfers. Every backtest reads the same store. Every researcher works from the same ground truth.
* Stylised visual. No representation of past, current or projected performance.
