Paper trading: the step between a backtest and a real decision

Not a demo button

The environment keeps real bookkeeping: cash, positions, average cost, commission paid, realised and unrealised profit, and a reconciliation against the latest price we hold. Fills are simulated at the reference price plus the same half-spread the backtest charges, so a strategy cannot look better in simulation than it did in testing by quietly getting a better price.

A signal at t cannot trade at t

The execution model is explicit: a signal dated on a given session is acted on at the next available session or later. If the newest signal is the newest bar we hold, nothing trades, and the system says why rather than filling at a price that was not knowable when the signal was formed. That single rule removes the most common way a research result flatters itself.

Four independent stops

1. A division-wide kill switch — stops every simulated strategy at once.
2. A per-account switch — halts one strategy without touching the others.
3. A per-instrument halt — blocks a single symbol across an account.
4. Closing-only mode — existing positions may be reduced, but no new exposure can be opened.

Alongside those: a cap on the size of any one order, a cap on total simulated activity per day, and a limit on how many positions can be open at once. Every order, accepted or refused, is written to an audit log with the reason.

Live trading is not switched off — it does not exist

There is no brokerage adapter in this platform. The order path refuses any mode other than simulation, so no configuration change can start real trading; it would take new code plus an explicit, signed authorisation under our highest governance tier, with independent risk limits and a kill switch defined in advance. A strategy may only enter paper trading as a challenger after a person approves it. Nothing is promoted automatically.

What paper trading still cannot tell you

• Whether your order would actually have been filled in a fast market.
• How much size the market could have absorbed at that price.
• Whether the next regime resembles the one being sampled.
• Anything about your own account. This is a research environment; it does not manage money and it makes no promise about any outcome.

Evidence and sources

1. 1. Kissell, The Science of Algorithmic Trading and Portfolio Management (2013) — Transaction-cost components we charge in simulation: commission, spread and market impact.
2. 2. Almgren & Chriss, Optimal Execution of Portfolio Transactions (2000) — Why a fill is not the closing price, and why capacity assumptions belong in a specification.
3. 3. Bailey & López de Prado, The Deflated Sharpe Ratio (2014) — Why a forward simulation is treated as one more test to be discounted, not as proof.