Reading a verdict¶

The idea¶

Arvo's answer to a study is not a number. It is a verdict, a paragraph of advice, and then the evidence — in that order, deliberately, because a number always looks like an answer.

Verdict	Means
Supported	beat the benchmark by the required 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 the question either way

The criteria are fixed and published rather than tuned per run: at least 30 trades, an excess return over the benchmark above 0, and a drawdown under 30%. The benchmark is holding the instrument, which is the bar every study is measured against — making money is not the test, beating the alternative of doing nothing is.

Why Inconclusive is the important one¶

Inconclusive is the most common verdict, and it is not a failure, a limitation or a soft NotSupported. It is a different kind of statement:

NotSupported says the rule did not clear the bar on this evidence.
Inconclusive says there is not enough evidence here to say either way.

A rule with 18 round trips has not been shown to work or to fail. Reporting a return on it would be reporting noise with a decimal point. Most systems in this space would give you the ratio anyway; Arvo says it does not know, which is the honest answer and the one most people are not used to hearing.

The correct response to Inconclusive is a better test, not a different rule. Widen the window, or pool evidence with a panel — the same configuration across many instruments, which makes the trade-count bar reachable honestly instead of by lowering it.

If you are getting Inconclusive constantly, nothing is wrong

A daily rule on one instrument over two years produces a few dozen trades at most. That is the arithmetic of daily bars, not a fault in your rule or in Arvo. The intraday families — mean reversion, opening range — generate enough round trips to get real verdicts, which makes them the better place to learn what a verdict feels like.

Advice, and why it comes first¶

Every finding carries advice: named items, each with a finding in words, its evidence, an action, and a severity.

Severity	Means
`blocking`	the numbers below are not a result yet
`warning`	read them with this in mind
`info`	worth knowing

Real examples Arvo produces:

Too few round trips to tell skill from luck — 18 trades against a 30 minimum — widen the window or loosen the entry; do not read the return until it does.
The run ended holding a position — treat the tail of the curve as provisional.
The point estimate says nothing about its own error — lengthen the window; do not compare this Sharpe against another until it is.
The selection did not survive deflation — the best of this many tries is what luck produces.
The dividend gap is most of the margin — the split-adjusted series cannot see distributions the benchmark paid.
Supported only under the stated costs — do not promote this; widen the edge or trade less often until the rule survives fills that cost twice what was assumed.

Each of those is a lesson from an earlier level, fired automatically against your own result. That is the whole design: the advice is the curriculum, applied.

The order to read a finding in¶

Follow this every time. The order exists because reading it in any other order makes you credulous.

The verdict. Then read_this_first, one sentence on what the verdict means for this run.
The advice. If anything is blocking, stop — the numbers below are not a result. Fix the test and re-run.
The search. How many configurations were tried, and whether the selection survived deflation against that many. A grid that only wins in sample is the most common way a wrong answer looks right.
Out-of-sample numbers only. The in-sample numbers chose the configuration. They are not evidence of anything and are shown so you can see the gap.
Against the benchmark, with the dividend gap beside it. If the gap is most of the margin, the margin is dividends the price series cannot see.
The cost tiers. Supported at stated costs but not at conservative means the edge belonged to the cost assumption.
The trades. Every round trip with the rule that fired, its value, the regime at entry, the exit reason, fees, and what filled against what was asked.

Cost tiers¶

Three, and only two of them can affect a verdict:

Tier	What it is	Can it change a verdict?
Realistic	the experiment's stated model: what a fill is expected to cost at that venue	it is the verdict
Conservative	half again the commission, twice the slippage and never under five basis points, twice the fees, twice an option's spread	it can only refuse, never upgrade
Optimistic	half the commission and nothing else	never — nothing is judged under it

Optimistic exists purely to answer "how much of this result is costs?" The leaderboard ranks by what survived conservative costs first, so a rule that looks better raw and worse at doubled costs ranks below its neighbour. Raw return is a column you read and never the order — a ranking by raw return is the overfitting machine every screener ships, and Arvo does not ship one.

A session's verdict is a different question¶

Once a rule is trading, the same three words are reused for a new question: do the live trades still look like the finding's out-of-sample trades?

Verdict	Means
Holding	the live trades sit inside what the out-of-sample distribution would produce
Diverging	expectancy, drawdown or firing rate has left that range; the reason is named
Inconclusive	too few live trades to say, which is where every session starts

The named reasons, with the thresholds, so a verdict can be read rather than trusted:

Reason	Fires when
`expectancy`	after ten closed live trades, the live mean is more than two standard errors below the finding's
`drawdown`	at one and a half times the finding's out-of-sample maximum, at any trade count
`frequency`	four entries were due and fewer than a quarter, or more than four times, as many came
`regime`	most entries, and at least three, landed in a regime the finding never traded in
`execution`	after five fills, slippage is past twice what the cost model assumed

A verdict is not a risk limit. Diverging changes nothing at the gate — it is recorded, raised as an alert, and is the case for demoting a session, decided by a person. The drawdown halt and the daily loss limit protect the account; the verdict judges the rule. Keeping those two jobs separate is deliberate.

Staleness¶

A finding carries stale_reason: data, ruleset, or both. The finding stays — it simply no longer describes the present. Re-run it; the old one remains as history.

A finding records the ruleset's content hash, the engine's commit and the dataset's version, which is what makes staleness detectable at all rather than something you have to remember.

What people get wrong¶

Reading the return first. The whole page is arranged to stop you doing this, and it is still the most common mistake.

Treating Supported as permission. It means one finding cleared one bar on one window. Real money needs the promotion gate, and that is level 9.

Comparing two findings' Sharpe ratios casually. If the advice says the point estimate says nothing about its own error, then the comparison is noise against noise.

Assuming a verdict is about the account. It is about the rule. The account has its own protections and they are unrelated.

Watch¶

(to be chosen) — statistical significance and sample size, in plain terms.
(to be chosen) — why "not enough evidence" differs from "no effect".

Next¶

Paper to real money — what paper trading measures, what it cannot, and the gate in front of the money.