Swing Scan

Build a swing trading entry rule, backtest it honestly, and compare exit strategies on identical entries. A $0/month portfolio project, built in about ten days by directing AI coding agents.

Live demo · 52/52 acceptance criteria green · no sign up

Not investment advice. The public demo runs on a generated market, not real prices, and the generated market has a planted edge, so any strategy will look good on it. Nothing here is a recommendation to buy or sell anything.

🎬 Demo GIF pending. Until then, the live demo is the real thing and the demo script below is the two minute tour.

The idea

Most backtesting tools let you tune an entry rule until the equity curve looks good, which is how you end up with a strategy that fits the past and nothing else. This one is built around three things that make it harder to fool yourself:

1. The exit lab holds entries fixed. Pick one entry rule, then compare up to six exit setups on the same trades. Any difference in the results is caused by the exit, and that is asserted by an oracle test, not assumed.
2. A random entry baseline runs beside every result. Random entries in the same market go through the same exits. The gap between your rule and that baseline is the part that is not just market drift. In the published research note the chosen strategy's out of sample expectancy quadrupled while its edge over random fell 78% — the baseline is what caught it.
3. Every look costs you something visible. A trial counter keyed on rule structure (so changing a number does not reset it) counts how many times you have looked, and warns at ten. In sample results are highlighted; out of sample sits beside them, to be read once.

Plus the boring honesty: expectancy per bar held next to plain expectancy, MAE and MFE guides from in sample winners only, distinct entry weeks as the real sample size rather than trade count, and an assumptions header on every result stating the fill model, slippage and cooldown.

Demo script (2 to 3 minutes)

1. 0:00 The public URL, with the "Synthetic market: not real prices" banner. Free vendor data cannot be redistributed, so this is a generated market with planted delistings.
2. 0:15 The builder opens on the Breakout template. Change the volume multiple 1.5 → 2.0, add rsi(14) < 75, and watch the URL and the JSON panel update as you type.
3. 0:45 Scan results: "new today" flags and the operand values per ticker, so you can see why each row matched.
4. 1:00 A single config backtest: the assumptions header, the equity curve against DEMO-INDEX with the in sample and out of sample split marked, and the trade list filtered to exit_reason = delisted.
5. 1:20 The exit lab on five configs. Same entries, only the exit changes. Read the edge vs random column, then expectancy per bar beside plain expectancy: the trailing stop looks best per trade and loses per bar held. Read the MAE guide: "90% of winners never went below −6%, so an 8% stop is loose." Glance at out of sample once.
6. 2:10 Change the breakout lookback 252 → 100 and re-run. The trial counter moves and the overfitting warning appears.

Architecture

flowchart LR
    subgraph Presentation
        W["apps/web<br/>Next.js, static export<br/>Vercel"]
        A["services/api<br/>FastAPI, ~320 lines<br/>Lambda container"]
    end
    subgraph Application
        U["engine.api<br/>scan, backtest"]
    end
    subgraph Domain
        D["engine/<br/>indicators · rules · exits<br/>sim · baseline · metrics"]
    end
    subgraph Infrastructure
        I["engine/data, engine/synthetic<br/>engine/live<br/>Parquet, Alpaca"]
    end
    W -->|"typed OpenAPI client"| A
    A --> U
    U --> D
    U --> I
    I --> D

Every arrow points inward, and nothing points back out: the domain knows nothing about the API, the web app, Parquet or Alpaca.

Clean Architecture, with the dependency rule enforced by a test rather than a convention: engine/tests/test_engine_boundary.py fails the build if the domain imports FastAPI, Starlette or uvicorn. The API is ~320 lines because every decision lives inward of it; routers assemble responses and nothing else.

Exit precedence lives in exactly one function, step(position, bar), and "never special case an exit outside it" is an enforced invariant. Both the portfolio day loop and the per trade exit lab loop call it, so the two modes cannot disagree about what an exit does.

Read more: spec 0001 (stack) · all 10 specs · ADR digest, 17 decisions tracked to outcomes · agent log: how it was built

Tech, honestly

		Why
TypeScript (strict) on Node 24, Next.js App Router	static export	No server needed, so Vercel Hobby hosts it free
Tailwind + shadcn/ui, TanStack Query/Table, Lightweight Charts		Tables and charts are most of the UI
Python 3.12 (uv), FastAPI + Pydantic v2	Lambda container, ap-southeast-1	Pydantic models are the contract, and generate the TypeScript client
Polars (indicators and rules), NumPy and plain Python loops (simulator)		Deliberately not Numba or vectorbt. Profiling decided this, not taste — see ADR-006 and 016
No database, no auth, no background jobs		10 to 20 MB of Parquet ships inside the image. State lives in the URL
pnpm + uv monorepo, one Makefile		make ci runs everything CI runs

Things it deliberately does not have: accounts, saved rules, alerts, watchlists, OR groups in rules, intraday data, fundamentals, short selling, parameter sweeps, walk forward, Monte Carlo, or hosted real data. Those are listed as non goals, not oversights.

$0/month, specifically

Piece	How it is free
Web app	Vercel Hobby, static export, deployed from Git
API	AWS Lambda always free tier. 2,048 MB, 30 s timeout, x86_64, one container image
Image registry	ECR, images tagged with the commit SHA
Data	No database. Synthetic Parquet is built into the image
CI/CD	GitHub Actions, free on a public repo. Deploys use OIDC, so no AWS keys are stored
Staying awake	One /api/v1/health call a week, so the image does not go Inactive and spoil the first visit after a quiet spell. Not a warm up ping — the web app does that itself on page load
Guard rails	A $1 budget alarm and a throttle alarm, both by email

Measured cold start for a visitor hitting an idle function: 2.66 s (warm: 0.07 s). The first call right after a deploy is slower and different, 9 to 14 s, because Lambda optimises each new image once; infra/README.md §6 keeps the two numbers apart so neither gets quoted as the other.

Limits of the research

All of these point the same way: results look better than reality.

1. Survivorship (live data): Alpaca's free tier returns no bars for delisted names. Companies that died are simply missing.
2. Index selection look ahead (live data): the universe is today's S&P 500. Many names were added because they rose. A 2018 breakout in a company that joined the index in 2023 is a trade nobody could have screened for. This hits breakout and momentum rules hardest, which is exactly what the templates are, and it is probably bigger than plain survivorship.
3. One regime out of sample: the out of sample window is mostly a bull market. A long only momentum rule can pass it without being robust.
4. Short history: few sharp drawdowns in the window.
5. Multiple testing: many entry tweaks times up to six exits. Every look at out of sample makes it less out of sample.
6. Trade mode ignores capital: no slot competition, no correlated drawdowns. Distinct entry weeks are the honest sample size.
7. The synthetic market's edge is planted. A documented momentum term. Templates look good there by construction. Synthetic results demonstrate the method and the engine's correctness, never a strategy.
8. Costs are simple: 10 bps slippage, no commission, no borrow, no taxes.

Therefore every finding here is a hypothesis. Confidence would need delisted inclusive data, walk forward testing, and a paper traded forward test. The research note states its own confidence level and names which of these limits bite it.

Data licence

• The public demo, the GIF, CI and every fixture use synthetic data only. No real prices are committed, hosted, or shown anywhere public.
• Live mode is local only. Alpaca's free plan is personal, non commercial use only, so live bars are loaded to a gitignored data/live/ on one machine and never published. The API refuses to start in live mode anywhere but 127.0.0.1, and refuses on Lambda or in CI whatever the binding says.
• Aggregate metrics from live data count as derived data and are never published. Research notes run on live data stay in a gitignored research/ folder; only synthetic notes are published.
• Enforced rather than promised: a data leak guard and gitleaks run pre commit and in CI, and data/ and research/ are gitignored.

Running it

make setup      # install from lockfiles (pnpm --frozen-lockfile, uv sync --frozen)
make hooks      # once per clone: install the pre-commit hooks
make data       # generate the seed 42 synthetic market into data/synthetic (gitignored)
make dev        # API on 127.0.0.1:8000 + web on :3000
make ci         # every check CI runs, in one command

Live mode, local only, needs Alpaca keys in your environment or a local .env:

make load-live  # Alpaca daily bars into data/live (gitignored)
make dev-live   # the same app against your live data, on 127.0.0.1 only

Tests

Suite	What it is
26 oracle tests (tests/oracle/)	Owner written correctness tests, protected by a pre commit guard and a CI label so no agent can edit them. They define correct behaviour for every exit type, loop parity, look ahead and MAE/MFE, and they were written before the engine
141 acceptance tests (tests/acceptance/)	Written from the criteria and contracts only. All 52 MUST criteria are required and green; test_traceability.py fails the build if the gate and the traceability matrix disagree
710 Python tests, ~450 web tests	Unit tests beside their packages, plus Vitest for the UI
Golden references (tests/golden/)	Naive loop implementations that the fast paths are checked against

make test · make test-acceptance · make test-oracle · make lint · make typecheck

Recording the demo

The GIF above is the one artifact that needs a human. Record the six steps of the demo script against make dev on the seed 42 market, at 375 px wide (every page is checked to work there), and save it to docs/assets/demo.gif. Keep it under about 10 MB so it loads on this page, and make sure the synthetic banner is legible in the first frame — it is the first thing the GIF has to say.

---

A learning and portfolio project with one real user, never sold. Why there is no monetisation.