Monolith Research
Research infrastructure.
A classified research workspace, a rigorous evaluation discipline, and the data and tools that make every figure reconstructible.
This page describes how Monolith Research is organised — the library that classifies the research tree, the families of strategies built within it, the methods that separate genuine edges from artefacts, and the data and tools behind them. Strategy logic stays private; the surrounding evidence is made legible.
Classified, versioned, and held to one bar.
The working research root is a classified, versioned library — not a loose archive. Old notebooks, live candidates, raw caches, and published summaries are separated by use and by evidence status, and each family is indexed by what it contains, what produced it, and whether it is still active.
Failed branches stay in the library, because a branch that was built, tested, and closed is evidence — the cost of re-running a dead idea without that record is exactly the cost of building it again. Two parts of that record are public: the Strategy Graveyard, where retired research is documented in full with the real numbers, and the open-source validation gate — Probabilistic and Deflated Sharpe, PBO by combinatorial cross-validation — that holds every candidate to the same bar.
A documented failure is a first-class result. Σ Research discipline
The research spans twelve families built across the firm’s history — from mean-reversion and trend to microstructure, the equity cross-section, alternative data, and machine learning. Each strategy below expands to a short description; the current headline work is a research book of funding-grade sleeves, set out on the home page.
Mean-reversion & statistical arbitrage
Cointegration-driven spreads that revert to a stable relationship.
OU statistical arbitrage Researched
An Ornstein–Uhlenbeck mean-reversion engine over 90+ cointegrated cross-asset spread pairs, selecting the tightest, fastest-reverting handful each cycle.
Read the full noteCross-asset OU multibot Researched
The OU framework extended with a cross-asset momentum overlay and an adaptive universe that re-ranks each rebalance.
Read the full noteClosed-end-fund discount reversion Researched
Closed-end funds drifting to an unusual discount or premium against the value of their holdings, traded market-neutral on the expectation the gap reverts.
Read the full noteDual-class share spread Researched
The price gap between two share classes of one issuer with identical cashflows, traded long-cheap / short-rich on convergence.
Read the full noteTrend & momentum
Persistence and continuation across assets, sessions, and the futures curve.
Cross-asset trend following Researched
Time-series momentum across FX, equity indices, and commodities, sized by volatility.
Read the full noteCarry & trend Researched
FX carry combined with a trend filter, with risk balanced across positions.
Read the full noteJump momentum Researched
Reaction and continuation around intraday gaps and price jumps.
Read the full noteFutures term structure Researched
Curve-shape and roll-yield signals on futures markets.
Read the full noteIntraday index systems
Opening-range and intraday structure on equity-index futures.
NAS100 opening-range breakout Live
An opening-range breakout and intraday-momentum hybrid on the NAS100, with strict session and risk controls.
Read the full noteNAS walk-forward study Researched
Extended walk-forward testing of intraday index variants under stressed costs.
Read the full noteCommodity systems
Regime-filtered systems on gold and commodity markets.
Gold H4 system Researched
A daily-to-H4 gold system: a grey-model forecast, a multi-factor regime filter, and a confirmation layer.
Read the full noteGold directional overlay Researched
A directional overlay tested on gold around a directional-change signal.
Read the full noteMacro-regime ensembles
Multi-sleeve systems that switch behaviour with the macro regime.
Macro-regime ensemble Researched
A daily-to-H4 ensemble that combines several sleeves — a quadrant regime read, a factor-trend sleeve, and more — by volatility parity.
Read the full noteCross-sectional equity & factors
Ranking the equity cross-section by momentum, value, low-vol, and learned alpha.
Equity cross-sectional momentum Researched
Ranking the equity cross-section by relative momentum, long the leaders against the laggards.
Read the full noteLow-volatility equity factor Researched
The low-volatility anomaly tested on a long-only equity cross-section.
Read the full noteEquity seasonality Researched
Calendar and turn-of-period effects in the equity cross-section.
Read the full noteCross-sectional value Researched
A value factor built from a cross-sectional ranking of fundamentals.
Read the full noteSymbolic-regression equity alpha Funding-grade
A genetic-programming alpha miner run on a survivorship-correct, point-in-time US-equity panel, with a pre-registered blind holdout.
Read the full noteMicrostructure & order flow
Signals that live close to the tape — order-flow imbalance and the limit-order book.
Order-flow imbalance & DeepLOB Researched
Order-flow imbalance and limit-order-book features, screened foresight-first against the cost bar.
Read the full noteOrder-flow imbalance v2 Researched
A rebuilt order-flow pipeline that screens the foresight ceiling before the model — futures, equity, and crypto arms.
Read the full noteFutures microstructure Researched
Intraday microstructure features on a real futures order book.
Read the full noteAlt-data & nowcasting
Turning non-price data — weather, trade flow, positioning — into forecasts.
Weather-to-gas nowcast Funding-grade
Population-weighted degree-days nowcasting US natural-gas storage, walk-forward out-of-sample R² 0.946.
Read the full noteShipping-manifest equity signal Funding-grade
Customs bill-of-lading import flow read as a leading cross-sectional equity signal.
Read the full noteCOT positioning Researched
CFTC Commitments-of-Traders positioning tested as a cross-sectional factor.
Read the full notePost-earnings drift Researched
Post-earnings-announcement drift evaluated at scale on a point-in-time panel.
Read the full noteSentiment & NLP
News and social tone as a macro signal or a risk-warning layer.
Macro news sentiment Researched
GDELT tone and FinBERT scoring evaluated as a macro signal and risk-warning layer.
Read the full notePolitical-sentiment overlay Researched
A sentiment overlay from social and news flow onto gold.
Read the full noteCrypto
Market-neutral cross-sectional work on the digital-asset tail.
Crypto small-alt tail (Tidewater) Researched
A market-neutral, cross-sectional ML long-short on the small-altcoin tail, with a pre-registered three-stage gate.
Read the full noteMachine learning & crowd platforms
Learned models, candidate-generation engines, and independently-scored records.
Numerai RESIDUNET-CR Live
A crowd-residual model for the Numerai tournament — scored independently, with a live record.
Read the full noteAlpha factory Researched
A systematic candidate-generation and validation framework for testing many mechanisms quickly.
Read the full noteGenetic multi-strategy optimisation Researched
A genetic algorithm searching for a robust multi-strategy allocation.
Read the full noteFX systems
Multi-pair foreign-exchange forecasting, sizing, and execution.
Forecast-to-fill Researched
Smoothing → confidence → volatility-targeting → fractional-Kelly sizing with ATR exits across FX pairs.
Read the full noteFX strength & trend Researched
Multi-pair foreign-exchange session-strength and trend research.
Read the full noteIn-sample design — entries, exits, and the hypothesis stated before the test.
Blind-fold validation and out-of-sample splits, with costs stressed beyond the realistic estimate.
Simulated execution under real friction, with no capital at risk.
Small size, rule-based limits, and daily reconciliation — entered only after the gate is cleared.
Each research family has its own burden of proof, matched to the character of the strategy. Intraday candidates are checked through rolling walk-forward folds — each fold blind to the next — before a final held-out test. Mean-reversion candidates are split into formation (60%), validation (20%), and final holdout (20%) with no re-optimisation allowed between stages. The signal must behave in the holdout under exactly the parameters chosen during validation, or it does not advance.
Transaction costs are stressed in every evaluation. Spread estimates are inflated by a fixed multiplier; slippage assumptions are added at the position level. A strategy that cannot survive costs at 1.5× the realistic estimate is not carried into paper observation, regardless of its raw Sharpe.
The evaluation sequence is enforced by design. No candidate moves from in-sample optimisation directly to a capital test. The path is: in-sample, validation, final holdout, paper observation, then small size. Each gate requires a written record — the record is what makes any later result interpretable.
Negative results are kept alongside positive ones. The library records more closed branches than active ones. That is not a failure of research; it is the expected output of a process that holds every candidate to the same standard.
Small size. Daily reconciliation. No override.
When a candidate clears the gate, it is taken to a capital test at deliberately small size. The point is not scale; it is to expose the research to the parts of reality a backtest cannot reach — real fills, real spreads, statements, overnight handling, and the behaviour of the operator under live conditions.
Sizing is rule-based. Each candidate carries a pre-declared sizing approach, a per-position notional cap, and a portfolio-level drawdown trigger. The daily loss limit is set at 1% of equity; the trailing drawdown trigger at 2%. When either is breached, new entries are halted and the cause is written down before trading resumes. There is no discretionary override.
Reconciliation runs against records, not memory. Every fill that enters the performance record is matched back to a statement line or a timestamped API response. The same execution and reconciliation infrastructure stays in place for the next candidate that clears the gate.
- Data vendors
- Market and fundamental data from WRDS / CRSP (US equities, point-in-time), the EIA (energy inventories), the CFTC (positioning), IBES (estimates), OptionMetrics (options), and broker market-data APIs for FX and index candles. Numerai supplies the obfuscated tournament data behind the live model.
- Alt-data explored
- Open-Meteo weather (degree-days nowcasting gas storage), customs bill-of-lading import flow (shipping into equities), GDELT news tone, and CFTC positioning — each screened for genuine, tradeable foresight before anything else.
- Language & core
- Python throughout — pandas, NumPy, SciPy and statsmodels for the numerical and econometric core (Engle–Granger, ADF, half-life), numba for JIT-compiled rolling metrics, scikit-learn and XGBoost for learned models, and Optuna for walk-forward search.
- Data & backtesting
- Parquet snapshots queried with DuckDB and Polars for large panels; walk-forward folds and 60/20/20 out-of-sample splits; costs stressed at 1.5× the realistic estimate; final holdouts with no re-optimisation.
- Validation gate
- An open-source gate — Probabilistic and Deflated Sharpe, PBO by combinatorial cross-validation, a minimum track-record length, and a cost-and-beta strip — applied before any result is trusted.
- Performance measurement
- A metrics library of 50+ measures: Sharpe, Sortino, Calmar, maximum drawdown, time underwater, half-Kelly, rolling Sharpe, and walk-forward stability. Every published figure is produced by it from a versioned data snapshot.
- Version control
- Git for all code; dated Parquet snapshots for data; named output files for results. Any figure that cannot be traced back to a notebook, a data snapshot, and a named output file is not published on this site.
- This site — method, infrastructure, and operating discipline.
- The News page — working papers, note excerpts, and selected research outcomes with full metric disclosure.
- The strategy families, data families, and the principles behind sizing and reporting.
- Selected outcomes, where they can be reconstructed from a versioned record and a dated data snapshot.
- Negative results — every closed branch is documented with the reason for closure.
- Strategy logic — the specific entry rules, exit conditions, and overlay filters that define each candidate.
- The identity of the top-ranked pairs selected by the OU sleeve.
- Specific sizing parameters, per-position caps, and account records.
- Individual signals, their timing, and any data that would allow a strategy to be reconstructed from the public record.
- Credentials, execution switches, and venue configuration.
The boundary above exists so that the public record can be honest about method, evidence, and infrastructure without disclosing the signal logic behind any candidate. What is published here can be audited; what is kept private is kept private for a clear reason.
Questions on data, evaluation methodology, or reporting are welcome.
bilal@monolithresearch.uk