# Research data — provenance, and what is real money

*(Checked into git deliberately: `data/` is gitignored working state, so this
would not survive there. A copy sits at `data/research/README.md` next to the
files themselves; this is the canonical one.)*

Research-only datasets. **Nothing here is this system's own trading**, none of
it is managed by any script in this repo, and none of it may be merged into an
account journal or shown in the alpaca-trader UI (dan's explicit instruction,
2026-08-27).

The repo now holds three separate sources of trade data. Confusing them would
be the worst kind of error — two are real money and one is not. Always state
which a number came from.

| Dataset | Real money? | Who placed it | Where |
|---|---|---|---|
| **Alpaca `paper-main` / `Paper02qsr`** | **No** — paper | This system, automatically | `data/<accountId>/trades.jsonl` |
| **Alpaca `live-1`** | **YES — real money** | This system, automatically | `data/live-1/trades.jsonl` |
| **T212 round trips** | **YES — real money** | **dan, by hand** | `data/research/t212-round-trips.jsonl` |
| **Mock orders** | **No** — simulated | yahoo-screener, signal-triggered | `data/research/mock-orders.jsonl` |

## t212-round-trips.jsonl

Source: `GET localhost:4173/api/orders/round-trips` (yahoo-screener → Trading 212).
Written by `scripts/backfill-t212-research.ts`.

- **Real money, manually placed by dan.** Tests dan's EXECUTION of QSR, not the
  signal itself — he used "slightly different signals but basically the same as
  newlyA", and may have bought a day or more after a signal fired.
- Scoped to **April 2026 onwards** — dan started QSR on T212 then. The feed also
  returns 75 older trips back to 2024-11 from a different, pre-QSR way of
  trading; those are excluded by default.
- **`pctReturn` is severely censored by the disposition effect** — winners get
  sold and appear here, losers get held and don't. Hence an 89.7% "win rate",
  mean +16.19% against a median +4.63%, holds up to 615 days. Prefer `alpha`
  (return vs buy-and-hold, which controls for market beta); treat any absolute
  return from this file as an upper bound.

## mock-orders.jsonl

Source: `GET localhost:4173/api/mock-orders` (yahoo-screener).
Written by `scripts/backfill-mock-orders-research.ts`.

- **Simulated. No broker ever saw these.** Never quote a figure from this file
  without saying so.
- **Signal-generated, not hand-placed** — so this tests the SIGNAL, where the
  T212 file tests dan's execution of it. That difference is the point of having
  both.
- Carries a `signal` field naming which signal fired (`setup`, `newlyA`,
  `triggered`), which is what makes "was newlyA right?" answerable at all —
  historical grade data does not exist beyond ~1 month, so it cannot be
  reconstructed any other way.
- Buys pair to sells via `pairedId`.
- **Most buys are still open** (259 of 330 at first snapshot). Scoring only
  closed pairs reproduces the censoring bias that produced three false findings
  on 2026-08-26 — fast winners close into the sample, laggards stay outside it.
  Use fixed-horizon forward returns computed from bars, for open and closed
  alike.

## Analysing any of this

Use `src/journal/clustering.ts`. These datasets have the same repeated-row
problem the Alpaca journal does: several orders share a ticker and a day, and
outcomes are near-identical within a symbol-day. Collapse to symbol-day or
session for significance, report the collapsed n, and never quote a raw row
count as a sample size.