Guide · Process
Paper trading is a useful mechanical rehearsal and a terrible predictor of results. It gets the arithmetic right and omits every friction that decides whether a small edge survives.
That is a real list and it is worth doing. What it is not is evidence that the strategy makes money.
| Friction | What paper assumes | What real money does |
|---|---|---|
| Slippage | Fills at the quoted price | Fills a few ticks worse, much worse on gaps and thin names |
| Partial fills | The whole order fills | Limit orders fill partially or not at all, changing the risk you actually took |
| Market impact | Your order does not exist | Size moves the book in illiquid names, and you trade against your own footprint |
| Short borrow | Shorts are always available | Locate may be unavailable, expensive, or recalled mid-trade |
| Halts and gaps | Orders behave normally | A stop through a gap fills far below the stop price, so the loss exceeds 1R |
| Fees and financing | Often ignored entirely | Commission, regulatory fees, and margin interest apply to every round trip |
| Queue position | Instant fill at the touch | A resting limit sits behind everyone who got there first, and only fills when the price is going against you |
| Emotion | Every signal is taken | You skip the uncomfortable ones, which are frequently the good ones |
The last row is the largest by a distance. A backtest and a paper account both take 100% of signals. In real money you will hesitate after two losses, size down after a drawdown, and take a discretionary exit at the worst moment. A strategy's realised expectancy is the strategy's expectancy multiplied by your compliance with it, and compliance is never 100%.
The queue-position row deserves a note too, because it produces a specific and nasty bias. A paper engine fills your resting buy limit the moment the price prints there. A real one fills it only if enough sellers arrive to clear the book ahead of you — which happens reliably when the market is about to keep falling, and unreliably when it is about to bounce. Paper gives you the good fills for free and charges nothing for the bad ones.
Run the arithmetic on the only setup my walk-forward testing did not eliminate. The anchored-VWAP reclaim came out at +0.117R over 4,933 trades on a 129-symbol, 10-year universe, 95% CI +0.057 to +0.174. It is not a validated edge: a risk-matched random entry captured +0.086R on the same data, so the signal's own contribution is about +0.030R with a confidence interval crossing zero. The roughly double figure I first reported, on barely a hundred hand-picked trades, failed my own adversarial re-test and is retired. Take the headline 0.117R anyway, as the friendliest possible case, and assume a 1R risk of $300 per trade on 100 shares of a $48 stock.
modelled edge = 0.117R × $300 = $35.10 per trade slippage 2 ticks × 2 sides on 100 shares = $4.00 commission + fees = $2.00 ─────── net edge = $29.10 per trade = 0.097R friction = $6.00 = 0.020R = 17% of the modelled edgeTwo ticks a side and two dollars of fees — a genuinely benign assumption — takes 17% of the edge before anything goes wrong. That version still survives, barely, at 0.097R. Now assume slippage of 0.05R per side rather than a few ticks, which is realistic on gappy names and small caps, and you lose 0.10R of the 0.117R headline — over 85% of the entire result, gone to execution quality alone.
And remember what the headline number actually was. If the signal's own contribution over a risk-matched random entry is about +0.030R, then $6.00 of friction on a $300 risk unit is 0.020R, which is two thirds of the residual. On that reading the setup does not survive its own costs. A small edge is not robust to sloppy fills, which is the practical reason execution discipline matters more than signal hunting.
Worth building once, for your own instrument and holding period, because the items people forget are the ones that dominate. All figures below are on the same 100 shares of a $48 stock with a $300 risk unit, so 1R is $3.00 per share.
| Cost | Illustrative amount | As R | Notes |
|---|---|---|---|
| Slippage, entry | 2 ticks × 100 sh = $2.00 | 0.007R | Worse on gaps, thin names, and market-on-open orders |
| Slippage, exit | 2 ticks × 100 sh = $2.00 | 0.007R | Stop-outs slip more than take-profits, because they fill into momentum |
| Commission | $0 to $1.00 per side | 0 to 0.003R | Zero at most US equity brokers, per-share at institutional ones |
| Regulatory fees | Cents to about $1.00 on the sale | under 0.004R | Sell-side only, rate set annually — check the current schedule |
| Margin financing | $4,800 at 8% for 5 days = $5.26 | 0.018R | Only if on margin. Scales with notional and days, not with risk |
| Short borrow | $4,800 at 25% for 5 days = $16.44 | 0.055R | Hard-to-borrow names can run far higher, and the rate changes daily |
| Total, long, cash, 5 days | $6.00 | 0.020R | The benign case in the formula above |
| Total, short, HTB, 5 days | $22.44 | 0.075R | 64% of a 0.117R headline edge, and more than double the residual |
The bottom row is the one that ends short-side systems. A borrow rate is not a fee you pay once, it is a rate you pay per day on notional, and it has no relationship to your risk sizing. Two things follow: never quote a strategy's expectancy without saying which side it trades, and never assume the short leg of a long-short idea costs the same as the long leg.
A promotion gate is a pre-committed, written rule that decides when a system is allowed to trade real money, evaluated by something other than your enthusiasm. Without one, the promotion decision gets made on a good week, which is precisely the wrong sample.
A gate needs three properties to be worth anything:
My execution engine refuses to connect a live broker until the paper record shows 60 closed paper trades with positive realised expectancy. It is a hard check in the promotion path, not a note in a document, and the current record against it is zero closed paper trades, so the live path stays shut. That is not modesty, it is the actual state: the gate has never been approached, let alone passed.
The two numbers are chosen for different reasons. Sixty closed trades is a compromise: it is not enough for statistical confidence, which needs a few hundred, but it is enough to expose an engine that double-fires orders, mishandles a halt, or quietly loses on fees. Requiring realised expectancy rather than modelled expectancy is the important half, because realised includes the fills you actually got.
Written as something you could implement, it is about ten lines:
def guard_promotion_gate(journal, env): closed = journal.closed_trades() n = len(closed) exp_r = sum(t.realised_r for t in closed) / n if n else 0.0 if n >= 60 and exp_r > 0: return ALLOW if env.get("OVERRIDE_PROMOTION_GATE") == "I_ACCEPT_UNVALIDATED_RISK": log.warning("promotion gate overridden: n=%d expectancy=%+.3fR", n, exp_r) return ALLOW raise PromotionBlocked( f"gate: {n}/60 closed trades, realised expectancy {exp_r:+.3f}R")Three design details in that block matter more than the threshold. It reads closed trades, because open positions have unrealised results you would be tempted to count. It divides by n rather than summing, so a run of tiny wins cannot substitute for a real sample. And it raises rather than returning a flag, so a caller that forgets to check the return value still cannot trade.
An override exists in the code because a gate with no escape hatch gets bypassed by editing the source, which is worse. It is named to be uncomfortable to type, it must be set as an explicit environment variable, and it logs a warning with the actual trade count and expectancy every time it is used. That is the design intent: make the unsafe path possible, obvious, and recorded, rather than impossible and therefore routed around.
Worth reporting because it is the kind of bug that renders a gate decorative. My promotion check reads a SQLite journal maintained by the execution engine. My screener's paper runner writes a CSV journal somewhere else. Nothing bridges them — so sixty closed trades in the screener's journal would have advanced the gate by exactly zero, while looking on a dashboard like progress.
I deliberately did not auto-bridge them. Feeding a real-money gate from a system it was never written to validate defeats the purpose of the gate: the point is to prove the executor that will actually place the orders, not some other program that produced similar-looking rows. The correct fix is to run the paper phase through the executor being validated. If your gate reads from a different source than your paper engine writes to, you do not have a gate. Check that specific thing before you trust yours.
A promotion gate is only as good as the numbers feeding it, and the numbers are wrong more often than people assume. Reconciliation means comparing, per order, what your system thinks happened against what the broker says happened.
Reconciliation is also what turns the friction table above from a guess into a measurement. After sixty paper trades you should be able to state your own median slippage per side in cents and in R, rather than assuming two ticks because a guide said so. The reading a backtest report guide covers the equivalent discipline for backtest output, and the journal guide covers the field list.
| Stage | Size | Sample | What it is testing | Advance when |
|---|---|---|---|---|
| 1. Paper | Normal risk, simulated | 60 closed trades | Plumbing, order lifecycle, journal accuracy | Gate passes on closed trades and the journal reconciles to broker records |
| 2. Live minimum | 1 share, or the smallest tradeable unit | 20 trades | Real fills, real fees, and your own behaviour. Not profitability | 20 trades closed with zero reconciliation errors |
| 3. Live quarter risk | 0.25% instead of 1% | 30 trades | Whether realised R tracks paper R per setup | Realised and paper expectancy agree within a stated margin |
| 4. Live half risk | 0.5% | 30 trades | Whether stage 3 held up as size began to matter | No regime of systematic slippage growth with size |
| 5. Full risk | The planned risk percentage | Ongoing | The actual strategy | Never stop measuring |
Stage 2 is the one people skip and it is the one that earns its keep. A single share costs almost nothing and tests everything that paper cannot: whether your order actually routes, whether the fee schedule is what you assumed, whether the broker's stop order behaves the way its documentation claims, and whether you personally can watch a real loss without touching anything. None of those are strategy questions and all of them can end a strategy.
Stage 4 exists because slippage is not constant in size. A one-share order fills at the touch; a five-hundred-share order in a thin name does not. If your realised slippage per side grows meaningfully between stage 3 and stage 4, you have found the capacity limit of the strategy, and that is a fact about the strategy rather than a fact about a bad week.
Promotion rules without demotion rules are a ratchet. Write both, before you start, and code both.
The drawdown number is not arbitrary either. Recovery is asymmetric — 10% down needs 11.1% back, 20% needs 25%, 50% needs 100% — and the curve steepens exactly where discipline is weakest. The drawdown math is here, and it is the reason a halt threshold should be well inside the drawdown you think you could tolerate.
A side-by-side on paper environments, API quality, and the terms-of-service question is in the broker comparison guide. The short version: official paper trading on the same API is worth more than commission savings, because it is the only thing that lets a promotion gate mean anything.
If any line in that list is unchecked, the honest answer is that you are not ready, and the gate exists precisely so that answer does not have to be re-litigated on a good morning. The measurement layer has to exist before any of it is meaningful, which means logging planned 1R, realised R, and setup tag on every trade from the first paper fill onward. Journal fields that make this work → · TradeLog does it offline and locally · why R rather than dollars.
Educational content, not financial advice. No live profit-and-loss figures are claimed anywhere on this site; backtest and walk-forward results are always labelled as such. Full terms: /terms.html
FAQ
Slippage, partial fills, market impact, queue position, short-borrow availability, halts and gap fills, commissions and financing, and above all emotion. Paper accounts and backtests take 100% of signals, while a real trader hesitates after losses and takes discretionary exits, so realised expectancy is always lower than modelled expectancy.
Yes, but for engineering reasons rather than for proof of profitability. It verifies that orders route correctly, brackets attach, positions reconcile, and signal frequency matches the backtest, and it catches bugs such as duplicate orders and timezone errors that a backtest cannot expose.
A promotion gate is a written, pre-committed rule that decides when a strategy or system may trade real money. To be useful it must be defined before the paper run begins, measured in closed trades rather than elapsed time, and enforced by code or a third party rather than by the trader's own judgement.
My own execution engine refuses to connect a live broker until there are 60 closed paper trades showing positive realised expectancy, and the current record against that gate is zero closed paper trades. Sixty is not statistically conclusive, which needs several hundred trades, but it is enough to expose engines that double-fire orders, mishandle halts, or lose quietly to fees.
Enough to decide it. A headline 0.117R walk-forward result is 35.10 dollars per trade at a 300 dollar risk unit, and two ticks a side plus two dollars of fees costs 6.00 dollars, leaving 29.10 dollars or 0.097R. That already takes 17% of the headline, and it matters more than it looks because a risk-matched random entry captured +0.086R on the same data, leaving the signal's own contribution at roughly +0.030R with a confidence interval crossing zero, so 0.020R of friction eats two thirds of the residual. Slippage of 0.05R per side instead removes 0.10R, which is over 85% of the headline, so execution quality is often larger than signal quality.
Slippage on both sides, commission, regulatory fees, margin financing if you use margin, and short borrow if you short. The last two are the ones people forget and they scale with notional and days rather than with risk. On 100 shares of a 48 dollar stock at a 300 dollar risk unit, a five-day long on cash costs about 0.020R while the same trade short in a hard-to-borrow name costs about 0.075R.
Pull fill records through the broker API rather than the web interface, match on your own client order ID rather than on symbol and time, then compare intended price against fill price to get realised slippage and intended quantity against filled quantity to catch partials. Compare your computed position against the broker's reported position every close, and alert on any unmatched order in either direction.
In stages rather than with a switch. Pass the gate on paper, then trade the minimum size for about 20 trades to test fills and behaviour, then quarter risk for 30 trades while comparing realised R against paper R per setup, then half risk for 30 more to see whether slippage grows with size, and only then full risk. Pair it with coded halt and demotion rules, such as stopping at a 10% drawdown.