Clover Capital

Technology

A strategy engine that evolves every day

Clover Capital runs on a proprietary systematic platform: a point-in-time data moat, a dynamic quant engine that refreshes and re-weights daily, fundamental fusion and manager review, an automated and auditable execution chain, and real-time risk control — all sharing one point-in-time core, so the engine only ever acts on what was truly visible on the decision day.

The integrity problem

Most backtests quietly cheat

Two failures account for a large share of research that looks brilliant on paper and disappoints in practice. Both come from using data that an investor could never actually have seen at the time. The platform is built to rule them out before an idea reaches the portfolio.

The trap

Look-ahead & survivorship bias

Look-ahead bias: a model is fed figures — restated earnings, a corrected price, a late filing — that were not yet known on the date being simulated. The backtest sees the future.

Survivorship bias: testing only against today’s index members silently deletes every company that was delisted, acquired or went bankrupt — so the universe is composed entirely of names that already won.

Our answer

Reconstruct what was knowable

Every fact in our database carries both the date it refers to and the date we learned it. Any backtest can ask the database to reconstruct the exact state of knowledge at a past instant — prices, fundamentals and index membership as they were genuinely believed then.

Ordinary backtests see the revised past; we replay, as-of, what was truly visible on the decision day. A flattering result is therefore evidence of an effect, not an artefact of hindsight.

Data moat & signal universe

Broad data sources, a deep signal library, all ingested point-in-time

Four classes of data feed the engine — market and reference, fundamental, holdings and analyst, and alternative — every one of them ingested as-of, so that 700+ signals across 62 families can be researched without hindsight.

point-in-time query

# what was knowable about AAPL on 1 Jan 2020,

# as we believed it at that instant?

as_of('AAPL', '2020-01-01')

bar as published then — no later restatements leaked in

Market & reference data

Prices, volume, indices, interest rates, FX, commodities and volatility — the market’s own record, stored as it was quoted on the day.

Fundamental data

Financial statements, earnings forecasts, valuation, cash flow, leverage and quality metrics — with restatements appended, never overwritten.

Holdings & analyst data

Institutional holdings, fund filings, analyst ratings, target prices and forecast revisions, timestamped to when they became public.

Alternative data

Satellite imagery, web traffic, card spending, supply-chain links, news sentiment and knowledge graphs — additional lenses on the same companies.

Point-in-time core

Every row records two clocks — the date the fact refers to and the date we learned it — so any as-of query reconstructs exactly what was knowable, survivorship-bias free.

Lineage & governance

Run-id lineage on every result, bounded forward-fill so stale values can never masquerade as fresh, and append-only storage that keeps history fully reproducible.

Strategy engine architecture

From data ingestion to portfolio: one testable, versioned, reproducible pipeline

Six stages carry every position from raw data to a monitored portfolio. Each stage is versioned and reproducible, and the whole pipeline evolves every day.

  1. Data sourcesmarket · fundamental · holdings & analyst · alternative
  2. Research coreclean & merge · feature engineering · signal library · scoring
  3. Dynamic quant enginedaily refresh · signal re-rank · model update · risk rebalance
  4. Fundamental fusion & reviewfundamentals weight signals · manager review
  5. Execution layerpaper trading · order execution · attribution · alerts
  6. Portfoliomonitored · attributed · evolving daily

Enforced throughout

Point-in-time slicing · bounded forward-fill · run-id lineage · walk-forward evaluation · daily evolution. These are properties of the pipeline, not habits of its authors.

Fundamental review

Financial quality, industry trend, valuation and event context filter and weight the quant signals; a fund manager with more than fourteen years’ experience then reviews algorithm-selected holdings against investment logic and risk controls.

One direction

Ingestion can write; research can only read, through a contract enforced at the database layer. Signals can never quietly mutate the record they depend on.

Systematic asset management

A dynamic strategy that evolves every day

Quant is not a static model but an engine that learns and updates its strategies every day. Markets change every day; so the strategy learns every day.

  1. 01

    Data refresh

    daily update

  2. 02

    Signal re-ranking

    re-rank

  3. 03

    Model learning

    weight update

  4. 04

    Holdings re-ranking

    risk adjustment

  5. 05

    Execution feedback

    re-optimisation

Updated daily

New data, new rankings, new weights — every trading day.

Self-evolving

Continuously adjusted as market conditions change.

Closed loop

Research → holdings → execution → feedback, and back to research.

Markets change every day; so the strategy learns every day.

The quant platform

Research, portfolio, execution and risk on one point-in-time core

On top of the data moat sits a multi-strategy platform that carries raw data through to monitored portfolios — paper-first by construction, with live brokerage connectivity guarded by multiple safety gates — and a quant research terminal that puts data, signals, risk and execution in one interface.

Multi-factor signal engine

700+ registered signals across 62 families — value, momentum, quality, growth, low-volatility, sentiment, ownership, event-driven, behavioural, alternative-data and AI signals — scored and aggregated daily across 9,000+ U.S.-listed stocks.

Walk-forward backtesting

A custom, point-in-time-correct walk-forward engine reports the deflated Sharpe ratio — alongside out-of-sample gates and random-portfolio placebo tests — explicitly penalising the multiple-testing and overfitting that make naive backtests look better than they are.

Portfolio optimisation

Mean-variance, minimum-variance, risk-parity, Black-Litterman, tracking-error and volatility-targeted construction, each driven by an underlying factor risk model rather than raw sample covariance alone.

Multi-layer risk control

Constraints on positions, sectors, turnover, factor exposure and liquidity; drawdown and volatility alerts; VaR / ES and stress testing; factor-based and Brinson attribution explaining where every result came from — risk is sized on purpose, not avoided.

Automated execution

A six-step chain — signal generation, portfolio instructions, pre-trade risk and compliance checks, smart order routing with execution algorithms, fill feedback, and post-trade attribution — running on unified data standards (FIX / FAST / JSON), low-latency infrastructure, strict permissions, end-to-end encryption and compliance record retention. Strategies run paper-first; nothing is promoted to live execution until it has earned it.

Quant research terminal & monitoring

Data, signals, risk and execution in one interface, sharing one point-in-time core: market overview, strategy list, factor-exposure heat maps, execution status, and an alert centre for risk, anomaly, signal and data events.

Research automation

A platform that searches for its own alphas

Beyond the hand-built signal library, the platform automates parts of the research process itself — always inside the same point-in-time framework, and always behind a human-reviewed gate.

700+
signals
62
signal families
9,000+
stocks covered

Automated alpha mining

Genetic programming and methods from the recent academic literature generate and score candidate alpha formulas across multiple quality dimensions, with a language-model judge assessing financial logic. Nothing reaches production without passing a gated, human-reviewed promotion step.

AI analyst agents

Fundamental, technical, sentiment and valuation analysts run in parallel and feed risk-manager and portfolio-manager verdicts. The agents live inside the same point-in-time framework as classical signals — their output is one more input to be validated, never an authority.

Bias testing, enforced in CI

Every registered signal must ship with its own lookahead-bias test, and a meta-test fails the build if any signal lacks one. Point-in-time correctness is a property of the codebase, not a habit of its authors.

Governance & rigour

Traceability, engineered

Traceability is one of our three investment rules, and the platform is built to honour it: every number in a research result can be followed back to its source, every trade to the signal and review that motivated it, and every result can be reproduced.

Lineage & auditability

Every fact is traceable to its ingestion run, vendor and raw source record; every result carries a run-id. Any number can be followed back to where it came from.

Validation on ingest

Incoming data is validated before it lands, with rejected records archived rather than silently dropped — bad data is visible, not invisible.

Read-only contract

The research platform’s access to the database is read-only and enforced at the database layer, so research can never corrupt the record it depends on.

Reproducibility

Append-only, bitemporal storage means a result computed today can be reproduced exactly, on the same state of knowledge, tomorrow or next year.

From engine to portfolio

See how the engine becomes a strategy

The platform exists to serve a method. See how Clover Capital combines quant breadth with fundamental depth in its investment approach and flagship strategy, or get in touch to go deeper.