Quantitative Researcher

Quantitative Researcher

When to Use

• Frame a research question, null hypotheses, and falsifiable claims before touching data
• Source, license-check, and profile market or macro datasets (missingness, staleness, corporate actions)
• Run descriptive and inferential statistics with documented assumptions
• Apply time-series or panel methods at workflow level (stationarity, autocorrelation, fixed effects—when appropriate)
• Design factor, signal, or alpha research with clear economic intuition and testable predictions
• Structure backtests with realistic costs, point-in-time universes, and bias checklists
• Compute and interpret risk metrics (volatility, drawdown, tail risk) and stress regimes
• Produce reproducible notebooks or research memos with limitations, sensitivity, and uncertainty bands
• Communicate what would change the conclusion—not point forecasts presented as advice

When NOT to Use

• Production ML pipelines, feature stores, model serving, or MLOps → data-scientist, ml-research-engineer-safeguards, ml-ops-engineer
• Executive BI dashboards, KPI definitions, or warehouse metric layers → data-analyst (if installed), bi-analyst, analytics-data-engineer
• Equity initiation, earnings narrative, or sell-side style research reports → equity-research, initiating-coverage, earnings-analysis (if installed)
• SOX close, journal entries, or GAAP financial statements → financial-statements, compute-accounting-manager
• Legal investment advice, suitability, or regulatory filings → commercial-counsel, compliance skills
• Trading execution, OMS, or low-latency production systems → senior-software-engineer (if installed)
• Product A/B tests and experiment platform design → ab-testing-engineer
• Text sentiment forecasting without quant factor/backtest framing → sentiment-forecasting-engineer, sentiment-analysis-engineer
• Alert threshold and false-positive decision policy → anti-false-positive-decision-making
• Bond RV, curve trades, or issuer credit narrative → bond-relative-value (if installed)
• Ratio analysis and corporate finance storytelling without research design → financial-analyst (if installed)

Related skills

Core Workflows

1. Frame the research question

1. State the decision or learning goal (not "find alpha" without a mechanism) 2. Define population, horizon, and frequency (daily bars vs intraday changes methods) 3. Pre-register primary statistic or metric; list secondary and robustness checks 4. Document null and alternative; specify what evidence would reject the hypothesis 5. List data requirements and known limitations upfront

See `references/research_framing_and_data_quality.md`.

2. Acquire and validate data

1. Record vendor, version, as-of rules, and adjustment policy (splits, dividends, total return) 2. Profile: coverage, gaps, duplicates, timezone alignment, survivorship in universe files 3. Run reconciliation spot checks against a second source where feasible 4. Freeze a research snapshot (hash, date range, universe version) before analysis

See `references/research_framing_and_data_quality.md`.

3. Explore and model (descriptive → inferential)

1. Start with descriptive stats and visual diagnostics (distributions, outliers, breaks) 2. Choose methods matched to dependence structure (i.i.d. vs time series vs panels) 3. Report effect sizes, confidence intervals, and assumption checks—not p-values alone 4. Run sensitivity to window, winsorization, and sample exclusions

See `references/statistics_time_series_and_panels.md`.

4. Factors, signals, and backtests

1. Tie each signal to an economic story and holding period 2. Build signals with point-in-time inputs only; document lag and publication delay 3. Backtest with transaction costs, capacity intuition, and turnover reporting 4. Audit lookahead, survivorship, selection, and overfitting (multiple testing)

See `references/factors_signals_and_backtesting.md`.

5. Risk, robustness, and regimes

1. Report volatility, drawdown, and tail metrics with window definitions 2. Interpret Sharpe and related ratios with known limitations (non-normality, short samples) 3. Segment by regime (vol, rates, liquidity) and run stress scenarios 4. Compare in-sample vs out-of-sample and walk-forward where applicable

See `references/risk_metrics_and_robustness.md`.

6. Deliver and document

1. Ship a reproducible artifact (notebook + pinned deps + data manifest) 2. Include limitations, assumptions, and uncertainty language suitable for stakeholders 3. Separate research findings from implementation or execution recommendations 4. Archive parameters, random seeds, and version metadata

See `references/research_deliverables_and_ethics.md`.

When to load references

Factors, Signals, and Backtesting

Table of contents

1. Economic intuition first 2. Signal construction 3. Portfolio formation 4. Backtest mechanics 5. Bias checklist 6. Overfitting and multiple testing 7. Performance attribution framing

Economic intuition first

A signal should answer:

• Why should this predict returns or risk? (risk premium, mispricing, structural flow)
• Who trades against it and why might the edge persist or decay?
• Horizon at which the mechanism operates
• Capacity intuition (liquidity, market cap segment)

Document failure modes (regime dependence, crowding, regulatory change).

Signal construction

Decay analysis: autocorrelation of signal and portfolio turnover vs horizon.

Portfolio formation

Common patterns:

• Sorts: quintiles/deciles on signal; long top, short bottom (or long-only top)
• Weighting: equal-weight vs cap-weight; impact on microcap bias
• Rebalance: calendar vs threshold; turnover implications
• Holding period: align with signal horizon; avoid overlapping return double-count without care

Report gross and net returns separately.

Backtest mechanics

Minimum backtest spec:

universe: [definition + PIT membership]
signal: [formula + lag]
rebalance: [frequency]
costs: [bps per side, spread assumption]
borrow: [short rebate / hard-to-borrow if applicable]
capital: [notional, leverage cap]
benchmark: [optional]

Include:

• Cumulative return and drawdown path
• Hit rate, mean spread, t-stat (with HAC where needed)
• Turnover and capacity proxy (ADV participation)
• Exposure to known factors (beta, size, value, momentum)

Bias checklist

Overfitting and multiple testing

Mitigations:

• Train / validation / test splits by time
• Walk-forward re-estimation of parameters
• Deflated Sharpe or haircut heuristics when many trials (cite method)
• Bonferroni / FDR across signal families explored
• Simple beats complex when performance is within noise

Prefer one primary backtest spec; others go to robustness appendix.

Performance attribution framing

Separate:

• Factor exposure (market, style) vs idiosyncratic residual
• Timing vs selection when relevant
• Costs drag vs gross alpha

Do not attribute in-sample fit as expected live performance without OOS evidence.

Quantitative Researcher — Scope

Table of contents

1. Role definition 2. In scope 3. Out of scope 4. Deliverables 5. Stakeholder interfaces 6. Quality bar

Role definition

The Quantitative Researcher supports markets and finance research where conclusions must be evidence-backed, reproducible, and explicit about uncertainty. The role spans question framing through data quality, statistical inference, factor/signal design, backtest hygiene, risk measurement, and research documentation.

This is not a production engineering role, not sell-side equity narrative writing, and not legal or personalized investment advice—though it informs decisions that others may implement under separate governance.

In scope

Out of scope

Deliverables

Typical artifacts:

1. Research memo — question, data, methods, results, limitations, falsifiers 2. Data quality report — coverage, adjustments, PIT checks, known gaps 3. Exploratory notebook — reproducible EDA with pinned dependencies 4. Signal spec — definition, lag, universe, neutralization, turnover expectation 5. Backtest summary — IS/OOS, costs, bias checklist, robustness grid 6. Risk dashboard (static) — vol, drawdown, regime splits, stress table 7. Sensitivity appendix — windows, winsorization, alternative universes

Stakeholder interfaces

Quality bar

Research is ready to share when:

• [ ] Research question and primary metric are stated before deep dives
• [ ] Data source, version, and adjustment policy are documented
• [ ] Point-in-time rules are explicit for signals and universes
• [ ] Assumptions for statistical methods are checked or flagged
• [ ] Backtests include costs, turnover, and a bias checklist
• [ ] Results include uncertainty (CIs, bands, or sensitivity)—not single-point certainty
• [ ] Limitations and what would change the conclusion are written
• [ ] Artifact is reproducible from manifest (deps + data snapshot reference)

Research Deliverables and Ethics

Table of contents

1. Deliverable types 2. Notebook and report structure 3. Reproducibility 4. Uncertainty communication 5. Research vs advice boundary 6. Data ethics and licensing 7. Peer review checklist

Deliverable types

Match detail level to decision stakes.

Notebook and report structure

Recommended memo flow:

1. Executive summary (5–10 bullets: finding, magnitude, uncertainty, caveats) 2. Research question and pre-registered primary metric 3. Data (sources, period, universe, known gaps) 4. Methods (assumptions, equations at high level) 5. Results (tables, CIs, key charts) 6. Robustness (grid summary) 7. Limitations and falsifiers 8. Appendix (extra specs, code pointers)

Notebooks: keep one logical path; move sweeps to appendix cells or separate notebooks.

Reproducibility

Minimum package:

• environment.yml or requirements.txt with pinned versions
• Random seeds for bootstrap and MC draws
• Data manifest (paths or DVC hashes; no secrets in repo)
• Config file for parameters (YAML/JSON), not hardcoded toggles
• Makefile or single entry command: make research

CI optional: smoke test that notebook executes on sample data.

Uncertainty communication

Use language that reflects epistemic limits:

Always include:

• Sample size and time span
• Sensitivity to key assumptions
• What would falsify the conclusion

Research vs advice boundary

This skill supports analytical research, not personalized investment recommendations.

• Do not tailor conclusions to an individual's financial situation
• Do not imply fiduciary or suitability determinations
• Flag when outputs require compliance or legal review before external distribution
• Separate historical simulation from forward promises

Stakeholders implement trades under their own governance; research informs, does not direct.

Data ethics and licensing

• Respect vendor ToS and MNPI walls (no material nonpublic information)
• Document personal data if alt data touches individuals (usually out of scope)
• Cite academic and vendor sources; do not plagiarize proprietary research
• Store credentials in secrets manager, not notebooks

Peer review checklist

Before circulation:

• [ ] Question and primary metric stated upfront
• [ ] PIT and survivorship addressed
• [ ] Costs and turnover included in backtests
• [ ] Multiplicity / overfitting acknowledged
• [ ] Risk metrics defined with windows
• [ ] Limitations and falsifiers present
• [ ] Reproducibility manifest attached
• [ ] Language avoids advice and certainty overclaim

Research Framing and Data Quality

Table of contents

1. Question framing 2. Hypothesis and pre-registration 3. Data sourcing 4. Corporate actions and returns 5. Point-in-time and survivorship 6. Profiling checklist 7. Research snapshots

Question framing

Start from a decision or learning objective, not from a dataset.

Capture:

• Asset class and investable universe definition
• Horizon (holding period, rebalance frequency)
• Sample period and rationale for start/end (IPO bias, regime change)
• Primary estimand (mean return spread, IR, hit rate, regression coefficient)

Hypothesis and pre-registration

Document before analysis:

1. Null hypothesis (e.g., no difference in mean spread after costs) 2. Alternative and direction if one-sided 3. Primary test statistic and significance or CI approach 4. Planned robustness (subperiods, alternative neutralization, alt data vendor) 5. Stopping rules for iterative research (avoid unbounded peeking without correction)

For exploratory work, label sections exploratory and cap claims accordingly.

Data sourcing

Maintain a data dictionary: field name, unit, frequency, source, transformation.

Corporate actions and returns

Define return construction explicitly:

• Price return vs total return (dividends reinvested)
• Split adjustment method (backward-adjusted prices vs separate factors)
• Currency conversion for ADRs and global universes (FX source and timing)

Reconcile a random sample of names across two sources after major corporate actions.

Point-in-time and survivorship

Point-in-time (PIT) means inputs available at decision time, not with future restatements.

Document universe entry/exit rules (IPO seasoning, liquidity filters, delisting handling).

Profiling checklist

Run on every new dataset:

• [ ] Row counts by date; detect gaps and duplicate keys
• [ ] Missingness heatmap by field and time
• [ ] Outlier scan (winsorization policy decided after viewing, not ad hoc per result)
• [ ] Cross-section size per date (survivorship drops show as jumps)
• [ ] Join integrity (signals ⊂ prices ⊂ universe)
• [ ] Label leakage scan (features correlated with future-only fields)

Research snapshots

Freeze artifacts for reproducibility:

snapshot_id: 2024Q4-us-equity-v3
data_vendor: [name]
pull_date: YYYY-MM-DD
universe_file: universe_2024Q4.csv (hash: …)
price_adjustment: total_return_backward
signal_code_git_sha: …

Store hashes or version IDs; avoid "latest" in final memos.

Risk Metrics and Robustness

Table of contents

1. Return and volatility 2. Drawdown and tail risk 3. Risk-adjusted ratios 4. Factor exposure and beta 5. Regime analysis 6. Stress testing 7. Robustness grids

Return and volatility

Define return frequency (daily, weekly) and annualization factor explicitly.

Report vol in percent and decimal consistently; label currency for multi-currency books.

Drawdown and tail risk

Show drawdown path chart with recovery time (time underwater).

Risk-adjusted ratios

When reporting Sharpe:

• State risk-free series used
• Prefer bootstrap CIs on Sharpe or returns
• Note non-normality and autocorrelation
• Avoid ranking strategies on in-sample Sharpe alone

Factor exposure and beta

Run regression of portfolio returns on factor returns (Fama–French, macro, custom).

Report:

• Betas + CIs (HAC SEs)
• R² and residual vol
• Rolling beta plots for stability

Distinguish risk exposure from claimed alpha after factor adjustment.

Regime analysis

Segment history by observable regimes:

For each segment, report mean return, vol, Sharpe, max DD, and N (effective sample).

Avoid too many segments on short histories (false precision).

Stress testing

Historical stress: replay portfolio through known windows (2008, 2020, 2022 rates, etc.)

Hypothetical shocks:

• Parallel rate shift (+100 bps)
• Equity shock (−20% beta-scaled)
• Vol spike (scale residuals)

Document linear vs nonlinear instrument assumptions (options, convexity).

Stress outputs are scenario illustrations, not forecasts.

Robustness grids

Standard grid for research memos:

Present as table of key metrics, not only best cell. Highlight sign flips and magnitude swings.

Conclude with stability score (qualitative): robust / mixed / fragile.

Statistics, Time Series, and Panels

Table of contents

1. Descriptive foundations 2. Inferential principles 3. Dependence and effective sample size 4. Time-series workflow (high level) 5. Panel data workflow (high level) 6. Common pitfalls 7. Reporting standards

Descriptive foundations

Before modeling, report:

• Central tendency and dispersion (mean, median, std, IQR)
• Skewness and excess kurtosis (fat tails matter for risk)
• Quantiles (1%, 5%, 95%, 99%) for return and signal distributions
• Rolling statistics with window length stated (e.g., 63-day vol)
• Correlation matrices with caution on synchronous overlap at high frequency

Visualize: histograms, QQ plots, rolling mean/vol, drawdown paths.

Inferential principles

Avoid interpreting p-values from many trials without multiplicity control.

Dependence and effective sample size

Financial series are rarely i.i.d.:

• Autocorrelation in returns and squared returns (vol clustering)
• Cross-sectional correlation on the same day (market factor)
• Overlapping samples in overlapping holding-period returns

Use Newey–West HAC standard errors for time-series regression, or block bootstrap for dependent data. Report effective N when using overlapping windows.

Time-series workflow (high level)

1. Plot levels and returns; identify structural breaks 2. Stationarity intuition: ADF/KPSS as diagnostics, not automatic truth 3. ACF/PACF for ARMA order hints; prefer parsimony 4. Seasonality for macro (monthly payroll, quarterly earnings seasons) 5. Regime splits when single global model is misleading 6. Out-of-sample evaluation with rolling origin or expanding window

Methods (workflow-level, not exhaustive implementation):

Panel data workflow (high level)

Panel = entities \(i\) over times \(t\).

1. Classify structure: balanced vs unbalanced; attrition reasons 2. Choose fixed effects (entity, time) when unobserved heterogeneity is plausible 3. Cluster standard errors at entity level (or two-way) for cross-sectional panels 4. Watch Nickell bias in short dynamic panels with lagged dependent variables 5. For factor portfolios, use Fama–MacBeth or panel regression with clear weighting

Common pitfalls

Reporting standards

Every inferential table should include:

• Sample period and N (entities × dates)
• Specification (FE, weights, SE type)
• Coefficient + CI (or SE) + economic interpretation
• Robustness pointer (alt spec in appendix)

Flag when conclusions are fragile to window choice or outlier treatment.