The first strategy I ever took live went from backtest to full size in one afternoon. The equity curve looked clean, so I saw no reason to wait. Within two weeks the live fills were so far from the simulation that I couldn't tell whether the strategy was broken or my code was.
I never found out which. That's the real cost of skipping steps: you lose the ability to diagnose anything.
A backtest is a claim
A backtest says: if the past repeats under these assumptions, here's what would have happened. Every assumption in that sentence can be wrong. Fill prices, fees, funding, data gaps, a bug that peeks one bar ahead.
So I treat the result as a claim to be tested, and I test it in four stages. Each stage answers one question, and I don't move up until it's answered.
Stage one: the backtest, attacked
Before it goes anywhere, I try to break it. Double the fees. Shift every entry by one bar. Run it on a period the parameters never saw.
If the result collapses when fees double, the edge was never there. It was a rounding error that happened to be positive. Better to learn that on a laptop than on an exchange.
Stage two: paper trading, for plumbing
Paper trading doesn't tell you whether the strategy makes money. Three weeks of fake fills can't do that. What it tells you is whether the machine does what you think it does.
Does it place the order when the signal fires? Does it handle a disconnect? Does it size correctly after a partial fill? I've watched bots paper trade perfectly for ten days and then double-enter on the first API timeout. Finding that costs nothing here.
Stage three: small size, for reality
Now real money, but a small amount. Think a tenth of the intended size, or less. The goal isn't profit. It's comparing live fills to the backtest, trade by trade.
Say your backtest assumes 0.05% slippage and the live average is 0.15%. On a strategy that makes 0.4% per trade, that gap eats a quarter of the edge. You can only see it if you're logging both numbers, and you only want to see it while the size is small enough to shrug off.
This is also where you find out how you behave. A strategy that's down 6% on a small account is a number. The same strategy down 6% on your full account is a different experience, and I'd rather know my reaction to it early.
Stage four: full size, in steps
Even after stage three works, I don't jump to full. I go to a third, then two thirds, then full, with a minimum number of trades at each step. The number depends on how often the strategy trades, but it has to be enough that one lucky week can't carry the decision.
Market impact also shows up here. Fills that were fine at small size can degrade as you grow, especially on thin pairs.
Why most people skip the middle
The common move is backtest straight to full size. It feels efficient, and I get the appeal. Waiting is boring, and a good-looking curve makes patience feel like a tax.
The other mistake runs the opposite way: paper trading forever. I think this one is quieter but just as costly. Paper results feel safe, so people keep collecting them instead of risking anything real, and the stage never ends because it can't produce the answer they're after.
A rule worth writing down
Before you start, write the exit criteria for each stage. For example: stage three passes if live slippage is within 30% of the backtest assumption over 50 trades. Fail it, and you go back, not forward.
Writing it first matters for the same reason a stop does. You'll be tempted to move the goalposts once real money is involved. At v33 Systematic we run every strategy up this ladder before it touches a client account, and the assumptions we test against are published at v33systematic.com.