Systematic Alpha Research Lab

Does this signal survive reality?

A focused quant research project that tests whether simple, explainable trading signals survive realistic validation — transaction costs, out-of-sample data, parameter changes, and market stress — on 20 years of real ETF data.

It is a Signal Survival Lab. Each signal is treated as a research case and walked through one question:

Pick a signal → understand the hypothesis → inspect the evidence
→ stress the assumptions → receive a verdict

The goal is not the best-looking backtest. The goal is disciplined research judgment: showing which signals deserve further research, which work only conditionally, and which should be rejected.

---

Research question

Do systematic alpha signals still look useful after costs, turnover, parameter changes, regime shifts, and out-of-sample testing?

A signal is credible only if it survives the whole process:

Hypothesis → Signal Rule → Portfolio Rule → Cost-Aware Backtest
→ Benchmark Comparison → Out-of-Sample Test → Robustness → Verdict

Signals tested

#	Signal	Idea (plain English)
1	Time-Series Momentum	Hold assets whose own 12-month trend is positive.
2	Cross-Sectional Momentum	Own only the strongest-trending assets relative to peers.
3	Short-Term Reversal	Buy assets that just sold off sharply, betting on a bounce.
4	Volatility-Scaled Momentum	Run momentum, but shrink positions when markets get volatile.
5	Equal-Weight Signal Ensemble	Average the four signals and hold what the committee likes.

Final verdicts

Verdicts are computed from the generated evidence using documented thresholds — never hand-assigned.

Signal	Verdict	Net CAGR	Sharpe	Max DD	Turnover/yr
Time-Series Momentum	✅ Survived	5.7%	0.57	−15%	3.2×
Cross-Sectional Momentum	✅ Survived	8.8%	0.62	−18%	4.2×
Equal-Weight Signal Ensemble	✅ Survived	7.8%	0.59	−22%	6.0×
Volatility-Scaled Momentum	⚠️ Conditional	4.3%	0.57	−12%	3.1×
Short-Term Reversal	❌ Rejected	5.1%	0.44	−29%	12.8×

Net of the primary 5 bps cost assumption. SPY buy-and-hold over the same window returned ~11% CAGR but with a −55% max drawdown — the strategies trade raw return for far steadier risk.

Headline: the momentum complex survives; volatility scaling is a risk-control overlay (conditional); short-term reversal is rejected — its 12.8×/yr turnover trails every passive benchmark and degrades sharply as costs rise.

Data

• Universe: 15 liquid ETFs across U.S. & international equity, Treasuries, TIPS, credit, commodities, gold, and a cash/T-bill proxy.
• Price sample: 2006-01-03 → 2026-08-14 on the SPY trading calendar (~5,186 trading days). Most ETFs cover the full span; DBC, SHV, HYG and BIL start later (see table) and each ETF enters the strategy on its own inception once it has the required signal history.
• Backtest calendar: strategies and benchmarks share a common calendar that begins 2007-02-28, when a 12-month momentum history first exists.
• Source: real Yahoo Finance prices using the dividend- and split-adjusted close (a total-return approximation), plus Yahoo macro series (VIX, 10-year and 13-week Treasury yields) and FRED CPI. NBER recession dates and fixed crisis windows are used only for retrospective regime labeling. No simulated prices are used anywhere.
• Cash proxy: cash/T-bill exposure (the cash fallback and the cash benchmark) earns the 13-week US Treasury-bill yield (Yahoo ^IRX), applied continuously. BIL and SHV are tradable ETFs, but the continuous T-bill rate is used for cash so there is no 2007 inception gap.
• Every series' provenance is recorded in research/outputs/data_audit/data_provenance.json.

ETF universe

Ticker	Asset group	First	Last	Source
SPY	US equity	2006-01	2026-08	Yahoo Finance
QQQ	US equity	2006-01	2026-08	Yahoo Finance
IWM	US equity	2006-01	2026-08	Yahoo Finance
EFA	Intl equity	2006-01	2026-08	Yahoo Finance
EEM	Intl equity	2006-01	2026-08	Yahoo Finance
SHY	Treasury	2006-01	2026-08	Yahoo Finance
IEF	Treasury	2006-01	2026-08	Yahoo Finance
TLT	Treasury (long)	2006-01	2026-08	Yahoo Finance
TIP	TIPS	2006-01	2026-08	Yahoo Finance
LQD	Credit	2006-01	2026-08	Yahoo Finance
HYG	Credit (high yield)	2007-04	2026-08	Yahoo Finance
DBC	Commodities	2006-02	2026-08	Yahoo Finance
GLD	Gold	2006-01	2026-08	Yahoo Finance
BIL	1–3M T-Bills	2007-05	2026-08	Yahoo Finance
SHV	Short Treasuries	2007-01	2026-08	Yahoo Finance

All 15 ETFs are included; none are excluded. The table is generated from data_quality_report.json (see research/outputs/reports/research_report.md).

Methodology highlights

• No look-ahead, enforced: feature[t] uses data through t; signal[t] is formed at t; position[t+1] uses signal[t]; return[t+1] is earned by position[t+1].
• Primary strategy: long-only, no leverage, no shorting, monthly rebalance, cash fallback when nothing qualifies.
• Turnover: sum of absolute weight changes at a rebalance — total traded notional / NAV, including both buys and sells.
• Costs: cost = turnover × bps / 10000, applied as a return drag. Reported at 1 / 5 / 10 / 25 bps; headline = 5 bps. Gross and net are always shown. Bid-ask spread, market impact, and intraday execution are not separately modeled.
• Benchmarks (same calendar & cost model): SPY buy-and-hold, 60/40, equal- weight universe, cash/T-bills, and inverse-volatility.
• Regime & crisis labels are retrospective validation tools (macro inputs lagged so labels use only observable information), not tradable real-time signals.
• Validation: train/test split (2006–2016 vs 2017+), expanding-window walk-forward with parameter selection, cost sensitivity, parameter robustness, rebalance sensitivity, six market-regime breakdowns, and five crisis windows.

Repository

research-lab/
  docs/                     # project specification (00–09)
  research/
    config/                 # universe / data / signal / backtest / validation YAML
    src/systematic_alpha_lab/
      data/                 # real price + macro ingestion, cleaning, returns
      signals/              # the five signal families
      portfolio/            # weighting, constraints, rebalancing
      backtest/             # engine, costs, metrics, benchmarks, runner
      validation/           # train/test, walk-forward, robustness, classification
      outputs/              # website JSON + research report generators
      pipeline.py           # end-to-end orchestrator
    tests/                  # return calc, signal lag, weights, costs, compounding, no-lookahead
    outputs/                # generated artifacts (audit / signals / backtests / validation / json / reports)
  web/                      # Next.js site; reads web/public/research-data/*.json

Reproduce

Research engine (Python 3.9+):

cd research
python3 -m pip install -r requirements.txt
PYTHONPATH=src python3 -m systematic_alpha_lab.pipeline   # fetch → signals → backtest → validate → JSON → report
PYTHONPATH=src python3 -m pytest tests/ -q                # correctness tests

The pipeline writes audit files, parquet outputs, website JSON (web/public/research-data/), and research/outputs/reports/research_report.md. Raw vendor responses are cached under research/data/raw/ so reruns are reproducible and offline. If Yahoo Finance is rate-limited on your machine, set TIINGO_API_TOKEN in .env for a keyed fallback — the pipeline never substitutes simulated prices.

Website (Node 18+):

cd web
npm install
npm run dev      # http://localhost:3000
npm run build    # static export to web/out/

The site reads only the generated JSON files. No research result, metric, chart, or verdict is hardcoded in the front end.

Limitations

• Long-only ETF research over one universe and window — not a live trading system and not investment advice.
• Transaction costs are a fixed-bps turnover drag; bid-ask spread, market impact, intraday execution, and taxes are not separately modeled.
• Adjusted-close returns approximate total return (dividends assumed reinvested without tax).
• Cash/T-bill exposure uses the 13-week Treasury-bill yield, not a cash ETF.
• Regime and crisis labels are retrospective validation tools, not tradable signals.
• Verdicts are qualitative research classifications, not statistical proof of alpha. Parameter grids are intentionally limited and predeclared to reduce data-mining risk. Past performance does not guarantee future results.

---

Research and educational project. Not investment advice. Not a live trading system.