// research lab
Where hypotheses fight for survival. Most ideas die on this board - and that's exactly how it should be. What survives, trades.
▌ 3 of 6 experiments currently live
▌
Automated exploration of network architectures for short-horizon signal extraction - letting the search space surprise us instead of hand-designing every model.
Hidden Markov cascades that classify the market's mood - trending, choppy, stressed - so every downstream strategy can adapt its behaviour in real time.
Translating order-book pressure into short-horizon directional forecasts, with careful attention to how market impact decays across venues.
Relative-value structures across the implied volatility surface. Paused while we study expiry-week liquidity before committing further capital.
Language-model pipelines that read filings, news and transcripts, and distil them into orthogonal factors a portfolio can actually use.
A reinforcement-learning agent that allocates risk across sub-strategies, trained against simulated regime shifts so it learns humility before it meets real markets.
// try the ideas
Two of the lab's core problems, turned into games. They're harder than they look - which is precisely the research finding.
A price feed will stream calmly - until, at a random moment, the regime breaks. Click the chart the instant you spot it. Too early counts as a false alarm.
the monitor is idle - start the feed when you're ready.
// research
Selected abstracts from the desk. The interesting parts stay in-house - but the ideas below show how we think.
★ featured
We explore quantum-inspired kernel methods for separating genuine structural breaks from noise in order-flow data, and discuss where these feature maps meaningfully outperform classical dimensionality reduction - and where they simply overfit beautifully.
abstract only · full paper on request
A study of queue dynamics in limit order books, examining when the shape and evolution of the book carries exploitable directional information - after realistic costs - and when it is merely noise wearing a convincing costume.
abstract only · full paper on request
Stacked hidden Markov models can flag transitions between market regimes earlier than simple volatility thresholds. We discuss the design choices that matter, and the failure modes that make naive implementations dangerous.
abstract only · full paper on request
Momentum in less liquid names is real but fragile: execution style determines whether the edge survives. We map how liquidity-aware scheduling changes the capacity and character of the strategy.
abstract only · full paper on request
Attention-based models that treat the implied-volatility surface as a sequence learn structure that parametric interpolation misses - particularly in the wings, where the interesting risk lives.
abstract only · full paper on request
Systematic option overlays can convert catastrophic drawdowns into survivable ones for a modest, budgeted cost. We frame hedging as insurance design and study what makes an overlay worth its premium.
abstract only · full paper on request