1. Data infrastructure
Quantitative research begins with trustworthy data. We ingest and normalise tick-level order-book data, trade prints, on-chain flows, derivatives positioning, funding rates, basis, and ancillary reference data across a curated universe of deeply liquid spot pairs on tier-one venues. All data are stored in a time-series warehouse with strict point-in-time discipline: every backtest and historical model evaluation uses only the information available at the corresponding historical moment.
Automated diagnostics test research-data quality. Gaps, outliers and venue outages are tagged, not silently imputed. Our research code treats missing data as first-class information rather than a nuisance to be papered over.
2. Signal generation
Our signal library spans three families:
- Factor models: cross-sectional and time-series factors rooted in market microstructure, carry, momentum, value, and liquidity. Each factor is specified ex ante with a clear economic thesis before any empirical test is run.
- Statistical learning: non-parametric and tree-based models constructed with leakage-free feature engineering, nested cross-validation, and explicit regularisation to control for the well-known pitfalls of machine learning in noisy, non-stationary markets.
- Microstructural signals: order-flow imbalance, queue dynamics, cross-venue dislocation and derivatives-spot basis, used primarily for short-horizon signal timing.
3. Validation discipline
A signal that looks good in-sample is not a signal. We impose:
- Out-of-sample segregation. A reserved period, never touched during development, is used as the final arbiter before any signal is retained as a research output.
- Combinatorial purged cross-validation. We adapt the Lopez de Prado methodology to address the serial correlation and overlapping-label issues that make naïve k-fold validation dangerously optimistic in financial data.
- Capacity stress tests. Each research model is evaluated under realistic slippage, participation-rate limits, and latency assumptions. A paper Sharpe that collapses under 5 bps of assumed slippage is discarded.
- Regime analysis. Performance is decomposed across volatility regimes, funding environments and correlation structures. We expect honest attribution, not a single number.
4. Model exposure design
For research purposes, signals are translated into hypothetical exposure profiles using optimisation methods that account for expected return, realised and forecast covariance, turnover, capacity and documented constraints. We prefer simple, explainable methods: every hypothetical model exposure must be attributable to an identifiable combination of signal, assumption and bound. Any actual position or order decision is made independently by the authorised manager.
5. Implementability
A research model that cannot be evaluated realistically is not useful. Each hypothetical exposure profile is assessed against slippage assumptions, participation-rate limits and capacity bounds so an authorised manager can perform its own independent assessment. DFI Labs does not execute orders, operate order-routing infrastructure or determine positions in client accounts.
6. Risk & governance
Risk diagnostics form part of model research and validation. During research validation, DFI Labs' tooling tests model outputs against documented research parameters; it does not monitor, supervise or control client portfolios or accounts. Any live pre-trade or post-trade risk control, order decision and supervisory response is carried out solely by the authorised manager.
Research governance is treated with the same seriousness as model risk: written methodology, documented decisions, continuity testing for our own research infrastructure and periodic review.
7. Research culture
Finally, methodology is only as good as the culture that sustains it. We run weekly research reviews, maintain a peer-critique requirement before any signal promotion, and track a living postmortem of decisions that did not work as expected, because improvement requires naming what broke. Small team, flat organisation, written thinking.
Want the detail behind the method?
We are happy to walk authorised asset managers and institutional counterparties through our research process under appropriate confidentiality. Conversations are best had on a call.