Paper to real money¶

The idea¶

A backtest assumes. A paper session measures. That difference is the entire point of this level, and it is why paper trading in Arvo is a measurement rather than a rehearsal.

Your backtest assumed a fill price, assumed every order filled, and assumed a latency of zero. A paper session finds out. What it finds out is called divergence: the measured difference between what the backtest assumed and what the venue actually did.

What paper trading measures¶

Measured	Against
Slippage — mean and worst adverse, in basis points across fills	the slippage the finding's cost model assumed
Latency — mean signal-to-fill time	the backtest's implicit zero
Unfilled orders — approved orders that never filled	a backtest, which assumes every order fills
Trade frequency	how often the rule fired in the window
Expectancy per trade	the out-of-sample trade distribution
Regimes entered	the regimes the finding's trades landed in

Unfilled orders are never averaged into the slippage figure, and that is the right choice: an order that did not fill is not a bad fill, it is a trade you did not get, and mixing the two hides both.

What paper trading cannot measure¶

Three things, and each one is a real gap rather than a rounding error.

Your own market impact. Paper fills do not consume liquidity. In a liquid large company at small size this is negligible. In anything thin, or at size, your real orders move the price you are trying to get, and no paper session will tell you by how much.

Queue position. A paper limit order at the bid is assumed to fill when the price trades there. A real one sits behind everyone who got there first, and on a fast move the price can trade through your level without ever reaching you.

What you will do. The hardest one. A paper drawdown and a real drawdown of the same percentage are not the same experience, and the difference decides whether you follow the rule. Nothing in Arvo can test this and no honest system claims to.

Say the cost

Paper trading closes the gap between a backtest and reality substantially, and it does not close it completely. Five days of paper is a filter against obvious breakage, not proof. Treat it as the cheapest available evidence rather than the last piece of it.

The promotion gate¶

Arvo does not let a finding reach real money because you feel ready. A session on alpaca-live or a Robinhood account is refused at start unless all three hold:

flowchart LR
    A["Verdict is<br/>**Supported**"] --> D{"Promotion<br/>gate"}
    B["**Five days** on<br/>alpaca-paper"] --> D
    C["That paper session<br/>**not Diverging**"] --> D
    D -->|all three| Y["Real money executor"]
    D -->|any missing| N["Refused, naming<br/>every reason at once"]

    style D fill:#1f6feb,stroke:#1f6feb,color:#fff
    style N fill:#b62324,stroke:#b62324,color:#fff

Details that matter:

The refusal names every reason at once, so they can be fixed together rather than discovered one at a time.
The gate is in the engine, on the start itself — so the command line, the window and an agent are all held to it. Nothing else creates a session.
Paper needs no promotion. The gate is only in front of real money.
The Sessions tab shows the road under the form as soon as a finding and an executor are chosen: the verdict, the days on paper, the paper session's verdict, and either the reasons it refuses or that it allows. The button reads Promote and stays disabled while the gate refuses.

What a live session does that a backtest does not¶

It audits its own book. Every poll, the session compares its book against the venue's. On a disagreement it freezes: it stops taking entries and waits for a person, rather than trading on a position it is not sure it holds.

A disagreement is not a bug, it is an event with real causes — a fill the process never heard, a hand-placed order, a broker-side liquidation. So:

State	Means
`running`	polling or streaming, taking entries
`frozen`	entries wait for a person; exits still go. Either the book disagrees with the venue, or the feed has gone dark
`halted`	the gate stopped the account: the drawdown halt, or the kill switch
`stopped`	asked to stop; positions left as they are

Reconcile makes the gate's book the venue's — the venue holds the money. Resume takes entries again and is refused until a reconcile has happened. Both are events on the record. Nothing resumes on its own.

It refuses yesterday's signals. A session started after the open catches up on the bars it missed: the rule is warmed on them and their signals go on the record, but an entry from a bar that closed before the session started is refused (catch-up: …) and never sent. Yesterday's signal at today's market is not the trade the finding measured. Exits still go.

It comes up halted if it inherits positions. A session starting against an account already holding positions adopts them and comes up halted, naming what it found, until you look. A limit that a restart lifts is not a limit.

The kill switch¶

Halt arms the gate and then flattens everything the session holds, in that order — so a proposer racing in is refused rather than opening into the exit. What the venue would not exit is named in the record and on the status, and the halt stands either way.

Available from a session's row, the Operations view, the tray (Halt trading, every session) and the command line:

arvo-engine ~/arvo session halt --all "spread blew out"

The drawdown halt is permanent for the run. The kill switch lifts only when a person releases it.

What people get wrong¶

Treating five days as a waiting period to be endured. It is a measurement. Read the divergence figures at the end of it — if your measured slippage is twice what the cost model assumed, the finding's edge may already be gone, and the execution divergence reason will say so.

Going live on a Diverging paper session by starting a fresh one. The gate reads the paper session's record, so this does not work, and it is worth knowing that the attempt was anticipated.

Confusing a freeze with a halt. A freeze is a session that is unsure of its position and wants a person. A halt is the gate protecting the account. Different causes, different fixes, and only one of them is about your money running out.

Sizing real money like paper money. The one thing paper cannot teach is what a drawdown feels like when it is real. Start smaller than the backtest justifies.

Try it¶

Take a finding with a Supported verdict and start a session on alpaca-paper. Paper needs no promotion, so this works immediately.
Let it run. Check session list or the Sessions tab for the divergence figures once anything has filled.
Compare measured slippage against what the cost model assumed. That single comparison is the most valuable number in this whole section.
Try to start the same finding on a real-money executor before five days have passed, and read the refusal. It will name every reason at once.

Watch¶

(to be chosen) — slippage and market impact demonstrated on real orders.
(to be chosen) — why paper trading results overstate live results.

Next¶

What moves the market — the last level, and the only one Arvo cannot test for you.