Research

You get an answer you can weigh, not a number you have to trust.

Run a study and Arvo comes back with a finding: a verdict, plain-language advice, and the evidence behind both. The verdict can be Supported, NotSupported, or Inconclusive — and that third one is the point. A rule with too few trades has not been shown to work or to fail, and Arvo says so instead of reporting a ratio that looks like an answer.

A study is the unit: one rule, one instrument, a search over the rule's parameters. It runs in the engine on bars from the project's library, and is stored as a finding you can reopen, replay, compare and export.

What happens when you run one

flowchart TD
    A["**Select**<br/><small>every point of the grid,<br/>on the first part of history</small>"]
    B["**Test**<br/><small>the winner only, on history<br/>it has never seen</small>"]
    C["**Discount the search**<br/><small>beat what the best of<br/>that many would score by luck</small>"]
    D["**Judge**<br/><small>30+ trades, beats buy-and-hold,<br/>drawdown under 30%</small>"]
    E["**Re-run at higher costs**<br/><small>can only refuse,<br/>never upgrade</small>"]
    F["**Record**<br/><small>verdict, advice, both curves,<br/>every trade, data fingerprint</small>"]

    A --> B --> C --> D --> E --> F
    style C fill:#1f6feb,stroke:#1f6feb,color:#fff

Each step in detail:

1. Selects on the first part of history. The rule runs over every point of the search grid, in sample.
2. Tests on history it never saw. Only the selected configuration, on the held-out window. These are the only numbers reported as the result.
3. Discounts the search. The best in-sample result is judged against what the best of that many random trials would have shown. A grid of forty tries that beats forty coin flips by nothing is nothing.
4. Judges it. Against the criteria — at least 30 trades, a positive excess return over buying and holding the instrument, a drawdown under 30%.
5. Checks whether the costs made it. Anything that would be Supported is run again at conservative costs: half again the commission, twice the slippage and never under five basis points, twice the fees, twice an option's spread. If it fails there the finding is refused, and the advice says it first — the edge was the cost assumption's, not the rule's. This step can only refuse, never upgrade.
6. Records everything. The verdict, the advice, both curves, the trade ledger with what the rule saw on each trade, the data's content hash, the ruleset's hash and the engine's commit — so a finding knows when it has gone stale because its data or its ruleset changed underneath it.

Kinds of study

Kind	Question
Study	Does this rule, over this search, beat holding the instrument out of sample?
Walk-forward	Does the selection hold up when the window rolls forward, re-selecting each step?
Panel	The same rule over many instruments at once, one parameter set: does it hold across names?
Book	Many instruments, one account, one gate: what does the portfolio do, and which members were silent?
Reported	Evidence a script computed elsewhere, submitted with its curve and ledger; Arvo computes the verdict, never the submitter

Verdicts

Supported — beat the benchmark by the margin, within the risk ceiling, on enough trades to be worth reading. NotSupported — ran cleanly and did not clear the bar. Inconclusive — the run cannot answer either way: too few trades, too short a window, a Sharpe whose standard error swallows it. The expected outcome, and not a failure. Verdicts and advice →

Every finding carries advice: named findings with a severity (blocking, warning, info), the evidence, and what to do about it — widen the window, loosen the entry, treat the tail as provisional. Read the advice before the return.

Where findings live

evidence/ in the project folder, one JSON per finding and an index. The History list in the Research view shows them with their staleness; the Runs tab shows the ones scripts and agents produced, grouped by author, with the size of the search each was held to.

Provenance

A finding records the ruleset's content hash, the engine's commit and the dataset's version. When the ruleset changes, every finding made with the old one is marked stale and says so; the same when the bars it read are refetched and differ. The version covers the span the run read, so a series that only gains new bars stales nothing.

The leaderboard

Every comparable finding in the store, in one order and only one:

1. Supported under the conservative cost tier, ordered by the mean profit per closed out-of-sample trade measured under that tier.
2. Supported under the stated costs where the conservative tier was never measured (findings older than the tier).
3. Everything else, whatever its number.

Ties break by drawdown, then by trades. Raw return is a column to read and never the order: a ranking by raw return is the overfitting machine every screener ships, and Arvo does not ship one. Each row carries the rule, the instrument, the interval, both verdicts, the expectancy and which costs it is under, the return, the drawdown, the trades, the regimes the trades opened in, the search size and whether the data has changed since.

It is the Leaderboard tab in the window (View, or the palette), where a row opens its finding and the rows stay in the order the engine ranked them. Sorting by a column there would be a second ranking. It is also the rank_findings tool for the agent, Research.RankFindings in the contract, and on the command line:

arvo-engine [<data-dir>] rank [--rule R] [--instrument I]

Panels are not rows: one configuration across many instruments is not the same number as one instrument, and the table would invite reading one against the other.