Field notes from the team building it.
Practitioner-grade write-ups on the decisions inside Nyquist — what we measured, what we cut, and where the boundary of what AI agents can actually do in production sits. No vendor demos, no hand-waved Sharpe figures.
The effectiveness of Lévy processes and stochastic-volatility models in pricing structured products.
Seven models calibrated to live July 2026 option chains on SPX, AAPL and SPY, then used to price four structured products. Smile fit runs from 6.45% IV-RMSE under Black-Scholes to 0.54% under Bates — and on the same market data the fair coupon on a barrier reverse convertible moves from 4.279% to 7.635%. The hedging-error tables say the left tail belongs to unhedgeable jump risk, not to model choice.
Correlation is a hyperbolic distance.
The parameter space of a stochastic-volatility model is not flat, and the cost of pretending otherwise is measured here: ρ = tanh(δ) to 1.6·10⁻¹⁵, Euclidean covariance averaging under-reports risk by 26–37%, straight-line stress-path interpolation distorts it by 27–108%, and calibration under the right metric converges in 2 iterations instead of 833. Every claim maps to a named table in the repository.
Hidden Markov model with Student's-t emissions for regimes of a bank bond portfolio return.
A four-regime multivariate HMM with heavy-tailed emissions on 904 trading days of a 124-issue RUB bond portfolio, validated the way a supervisor would. Over 403 rolling forecasts the Gaussian model breached its 99% VaR 10 times against a nominal 1% and was rejected by Kupiec and Christoffersen; the Student-t HMM breached twice and was not. The stress signal reaches an out-of-sample AUC of 0.88.
Unified quantum hydrodynamics of helium II: systematic exclusion of theory classes for the sign of transverse mutual friction.
A pre-registered exclusion program for a controversy open since 1965. Four of five reactive theory classes are ruled out against criteria fixed in advance; every formulation tested collapses to one algebraic boundary; the algebraic core is machine-checked in Lean 4 — 55 theorems, zero sorry, and the proof assistant found a sign error in the original analytics. Not a market model: what transfers is the method.
Green finance, ESG & sustainable finance: a literature review and research agenda.
A systematic review of 2018–2026 across five domains — green debt & the greenium, ESG rating divergence, climate risk as a priced factor, prudential stress testing, and the disclosure/taxonomy stack. Sustainable finance is now mainstream in form; a persistent measurement crisis and a political backlash leave it contested in substance.
Digital assets, Web3, DeFi & CBDCs: a literature review and research agenda.
A systematic review across five domains — AMM market economics, the DeFi risk taxonomy, RWA tokenisation, CBDC design, and the regulatory architecture. The ecosystem crossed from speculation to institutional infrastructure; the gap between innovation pace and risk-framework maturity is the systemic story of 2026–2030.
AI & machine learning in finance: a literature review and research agenda.
A systematic read of 2019–2026 across asset pricing, microstructure & RL, credit & fairness, NLP/LLMs and systemic risk. ML roughly doubles linear out-of-sample R², LLMs rival sell-side analysts, GNNs deliver a new tier of early warning — with opacity and strategy convergence as the unresolved cost.
Market digital twin: what urban planning taught us about modeling markets.
Singapore built a living replica of a city and tested interventions inside it before touching the real one. The five-component architecture of a market digital twin — real-time state sync, heterogeneous agents, an LLM semantic engine, a counterfactual engine, continuous recalibration.
Hybrid quant architecture: why neither pure AI nor pure classical works in production.
A London fund swapped its factor stack for a GPT-5.5 agent — Sharpe collapsed 1.4 → 0.6 in eight weeks. The five-layer hybrid stack that wins, plus the operational invariant: LLMs never do arithmetic that affects a trade.
R-Quant: the new profession AI just created in finance.
Statistician → Programmer-Quant → ML-Quant → R-Quant. The fourth speciation event in quant finance — a practitioner who orchestrates, evaluates and governs fleets of AI agents instead of writing one model at a time.
Why even Claude Opus 4.7 and GPT-5.5 fail at quantitative finance.
Five structural failures across every frontier model, and the 36-agent + NVIDIA NIM architecture Nyquist productionised at .09M/yr — a 70–83% cost cut vs. custom GPU clusters.
Your backtest is lying to you — the hidden cost of dirty market data.
Eight market-data failure modes that inflate apparent Sharpe by 0.2–0.5, and a worked example that walks Sharpe 1.61 → 1.19 under honest cleaning.
Quantum supremacy in finance — marketing or reality?
A line-by-line audit of the four canonical claims, and the one quantum task that already deserves budget today: NIST post-quantum migration.
No, you didn't build the next Bloomberg Terminal with Claude in a weekend.
The $200M data-licensing tax, the 18-month compliance onboarding, point-in-time correctness, and a framework for reading the next 'Bloomberg killer' headline.
What it actually takes to run 36 AI agents in production.
Three compute tiers, four architecture layers, six cost tables, the hidden clearing/connectivity/compliance bill — 18–24 months and $300K before a single line of alpha ships.
CeDeFi is the only realistic path to institutional DeFi adoption.
The four-layer architecture, five DeFi structural blockers, four CeFi weaknesses, and the quantitative risk frameworks that make it auditable.
Why finance still doesn't have a proper data layer — and what an ontology fixes.
The five semantic failure modes that consume 60–80% of every quant team's time, and the five-layer architecture that eliminates 80–90% of them.
AI Agents in Trading: where the real boundary of applicability lies.
Where LLM-based signal generation adds edge, where it destructively competes with classical factor methods, and what breaks at 9:31 AM on a CPI print day.
The same research set as a print edition — A4, with an abstract, keywords and an issue number. Made to be read on paper, or sent as one file.
From the write-ups to the live desk.
Part of this research is wired into the platform; the rest is context. Open the demo and watch the agent swarm debate on your own book — no calls, no scheduling.