Research charter

Mission: find investment strategies that are systematic (explicit rules, no discretion at run time), repeatable (same data → same decisions), and real (they survive honest out-of-sample testing, costs and different market regimes). Think like both a fundamental hedge fund (deep research into what drives a business or an economy, using unusual data) and a quant fund (turning that insight into rules, testing them ruthlessly, and combining many small edges).

1. Inspirations (what we take from each — not copies)

2. Data doctrine

1. Provenance or it doesn't exist. Every dataset has a manifest entry: source, URL, fetch time, coverage, checksum and quality findings (data/manifest.json).
2. Point-in-time only. A value is usable only after it was public. Every economic series carries a publication lag (claims +6 days, CPI +45, rail carloads +95). The engine applies it automatically.
3. Know your revisions. Revised series (payrolls, GDP, claims) look cleaner in hindsight than they were live. They are flagged. Upgrade path: ALFRED vintages.
4. Coverage matters as much as accuracy. A series that starts in 2023 silently disables any rule that relies on it before then. (Found on day one: FRED's ICE credit spreads only go back 3 years; we use Moody's Baa spread instead.)
5. Survivorship. Testing today's index members over history ignores the failures. Until we have point-in-time constituents, single-stock research is limited to clearly stated, survivorship-aware designs.
6. Alternative data is a research program, not a shortcut. Each new source (positioning, insider trades, attention, physical activity, imagery) needs the same treatment: lag, revisions, coverage, and a hypothesis for why it should predict returns. See python -m amberlab catalog for the roadmap.

3. Research protocol (enforced in code)

• Periods: train 2007–2015 · validation 2016–2021 · holdout 2022–today, sealed.
• Every evaluation is a logged trial (research/ledger.jsonl). The number of attempts per idea ("family") feeds the deflated Sharpe ratio, which discounts results for how hard we searched.
• Look-ahead detector: every evaluation recomputes the strategy on truncated history; any difference means it peeked at the future, and it is rejected.
• Stability over peaks: report parameter grids. Accept broad plateaus, reject lone spikes.
• Walk-forward: judge the selection process out of sample, not the best backtest.
• The holdout is opened once per final strategy, with a written reason, and the unsealing is logged.
• Costs always on: trading costs, financing on leverage and borrow on shorts.
• Realistic bar: after costs, a Sharpe of 0.5–1.0 is good. Above 2 on daily data means a bug or overfitting until proven otherwise.
• Paper-trade before capital: strategies that pass go to live paper trading (e.g. as a sleeve in Amber Terminal) for months, with pre-agreed shutdown rules.

4. Every strategy must state

1. Hypothesis: the economic or behavioral reason it should work, and who is on the other side of the trade.
2. Data: sources, lags and coverage.
3. Rules: exact, with parameters.
4. Expected failure modes: when it should lose money.
5. Results: train, validation, stability grid, walk-forward, deflated Sharpe, correlation to SPY and 60/40, and costs.

5. What's feasible here (and what isn't)

• Strong fit: daily and weekly multi-asset systematic strategies (trend, rotation, carry proxies, risk-based allocation, macro-regime models), economic and alternative-data signals, combining strategies into a portfolio.
• Later, with paid data: options strategies (needs historical option chains), single-stock factors (needs point-in-time fundamentals and constituents), intraday (needs minute data with history).
• Out of scope: high-frequency trading (requires tick data, co-location and microsecond execution).

6. Roadmap

1. ✅ Milestone 1: clean data layer, honest engine, research protocol, baselines.
2. ✅ Milestone 2: agent research team, using only the lab's tools (agents never edit the engine or see the holdout):
3. Researcher: proposes hypotheses from the literature and the data catalog.
4. Data scientist: profiles a new dataset (coverage, lag, revisions, predictive power).
5. Engineer: writes the strategy file.
6. Skeptic: attacks every result (leakage, overfitting, cost sensitivity, regime dependence).
7. Portfolio constructor: combines survivors.
8. Milestone 3: alternative-data integrations (CFTC positioning, SEC insider buying, EIA energy, Wikipedia attention), each as a profiled dataset.
9. Milestone 4: a strategy portfolio (pods with risk budgets), paper-traded through Amber Terminal.