Trading concepts

The vocabulary Arvo uses, and what each word means here — several of these terms are used loosely elsewhere, and the difference matters when you read a verdict.

For one-line definitions see the glossary. This page explains why each concept exists.

Rule, ruleset, strategy

Word	Means
Rule	One set of conditions: when to enter, when to exit. sma_cross is a rule
Ruleset	A rule plus a grid of parameter values to try. One ruleset is many configurations
Configuration	One point in that grid — fast: 10, slow: 30
Strategy	Avoided in the interface. It appears in code, the proto contract and findings for historical reasons

The distinction is load-bearing: a ruleset is a search, and how large the search was determines the bar the winner must clear.

Rules are data — JSON files under rules/, using TA-Lib indicator names and JSON Logic conditions — so a rule can be written, shared, diffed and replayed without rebuilding the engine.

Experiment, finding, evidence

flowchart LR
    R["Ruleset<br/><small>a rule + a grid</small>"] --> E["Experiment<br/><small>one configuration,<br/>one instrument, one window</small>"]
    E --> A["Artifact<br/><small>equity curve, ledger, refusals</small>"]
    A --> EV["Evidence<br/><small>metrics, advice, verdict</small>"]
    EV --> F["Finding<br/><small>stored under evidence/</small>"]

An experiment is a full specification: instrument, window, interval, which library version, which rule with which parameters, which costs, which risk model, which random seed. Everything needed to run it again.

A finding is what got stored: the evidence, the verdict, the advice, and the experiment that produced it. That is what makes a finding replayable — re-run it and the engine reports Reproduced or names the divergence.

In-sample and out-of-sample

The window is split. Configurations are chosen on the first part in sample, and the chosen one is judged on the second part out of sample, which it has never seen.

Choosing and judging on the same data is the single most common way a backtest lies. Arvo will not do it — and since the split is recorded in the finding, you can check where it fell.

Deflation

The correction for having searched. Given how many configurations were tried, what would the best of that many score if the rule had no edge at all? That number is the bar, and the winner must clear it.

Reported on every finding as the expected best under the null, alongside whether it was cleared. A result that beats a single-guess bar but not the best-of-nine bar is not evidence.

Deflation cannot see every bias

It corrects for searching. It cannot see survivorship bias — a universe of today's index members contains only companies that survived, and no search was performed to select them, so every gate passes such a result honestly. Knowing what a defence does not cover is part of using it.

Walk-forward

Instead of splitting once, re-select on a rolling schedule: choose on a window, trade the next one, roll forward, repeat. Answers a different question from a single split — would this have kept working as it was re-tuned over time — rather than did the one chosen configuration hold up.

Conservative costs

Every finding is re-run with costs increased. A rule whose edge is thinner than its transaction costs is common and looks fine at optimistic 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.

Regime

What kind of market a result was earned in — trending up, trending down, ranging — labelled after the fact over the closes.

Recorded on every trade. It lets you ask "does this only work in one regime?" rather than only "did it make money?", which is the difference between an edge and a bet on conditions continuing.

Regime labels are not causal

They are computed over a completed run, so they describe the past with hindsight. A rule may filter on a signal series only when that series is marked causal — computed from data available at that bar. The distinction is enforced in the data layer, not left to discipline.

The risk gate

Every proposed entry passes one gate: it is sized against the account, checked against limits, and either sized down or refused with a reason. The same gate runs in the backtest and live.

An exit never asks the gate — nothing may stop you shedding risk.

A refusal is never silent. It is on the record with its reason, so "why didn't it take that trade?" is answerable.

Session, executor, divergence

A session is one finding's rule running against one venue, keyed <finding>@<executor>.

An executor is where orders go: alpaca-paper, alpaca-live, robinhood-<last4>.

Divergence is the measured difference between what the backtest assumed and what the venue did — fill prices against decision prices, orders that never filled. It is why paper trading is a measurement rather than a rehearsal.

A session freezes when its book disagrees with the venue: it stops taking entries and waits for a person, rather than trading on a position it is not sure it holds.

Panel

One rule across many instruments, with one configuration chosen for the whole panel — not one per instrument.

Tuning per instrument is a second search: nine configurations over ten instruments is ninety chances to find something that fits. A panel also pools evidence, so the trade-count bar becomes reachable honestly instead of by lowering it, and it makes consistency visible — beating the benchmark on eight of ten is a different claim from beating it on one by a mile.

A panel is not a portfolio backtest

Nothing allocates capital across instruments or models the correlation between them. Each member is run with the whole account behind it, so the pooled drawdown is an average of separate drawdowns and understates what a real combined position would have suffered. Named rather than quietly averaged.

Universe

A named list of instruments with a required stated reason. A file without one is refused.

A list chosen by looking at what performed well is a search, and an unstated one is a search nothing counts. Requiring the reason in writing is the cheapest available defence.

Promotion

The gate in front of real money: a live executor is refused unless the same finding has run on paper for at least five days without diverging from its backtest. Days the session took bars on, not days on the calendar.

In the engine, so the command line, the window and an agent are all held to it.

Next

What Arvo does differently for why these exist, or the engine to use them.