The curve was beautiful until real fills got involved

The test was fitted to its own sample. The test skipped the costs you actually pay. You took trades the rules never described. And real money changed how you pressed the buttons.

Those four cover the great majority of cases, and they fail in different ways. Cost gaps make every trade slightly worse in a uniform way. Overfitting makes the live results look random rather than merely smaller than the test. Drift shows up as trades in your statement that have no matching signal in the test.

Start by measuring rather than guessing. Line the two records up trade by trade. The shape of the damage points at one of the four. Traders usually assume it is psychology, and the arithmetic usually finds costs first.

Most tests fill at the mid, instantly, at every price they asked for. Live, you cross the spread going in and again coming out.

Run the numbers on the E-mini S&P, where one tick is 12.50 dollars. Give up a single tick each way and that is 25 dollars gone. Say your commission and fees come to 4 dollars round turn. Every trade now starts 29 dollars behind where the test started it.

If the backtest showed an average of 45 dollars a trade over 400 trades, that is 18,000 dollars gross. Take off 29 dollars times 400, which is 11,600. You are left with 6,400 and one bad month wipes it.

Any system whose average trade is a handful of ticks is mainly a costs experiment. Re-run the test using your worst realistic fill.

Every time you nudge a parameter and keep the better result, you fit the test harder to the noise. That noise belongs to one stretch of data and nothing else.

Two checks catch it. First, hold back data you never looked at while building, then run the final rules on that block exactly once. A second pass turns it into more of the same sample. Second, move each parameter a few notches either side of your chosen value.

That second check matters more than people expect. A real edge sits on a plateau: a 12-period setting and a 14-period setting both work, roughly as well. A fitted result sits on a spike, where 13 prints a lovely curve and 12 and 14 both lose. If your settings are a spike, live trading will find that out for you.

Print the rules on one page before you start. Entry condition, stop, target, size, and the trades you are forbidden to take.

Then run it at the smallest size your broker allows. Fix the number of trades before the first order. Small size is the point. It keeps the money quiet enough for you to follow the rules while still making the fills real.

Keep a deviation log next to the trade list. One line per exception: signal skipped, entry chased, stop moved, or a trade taken with no signal behind it. That log is the measurement of the third and fourth gaps, and no amount of reflection replaces it.

At the end of the block, count deviations first. If a fifth of your trades were off-plan, the strategy has not been tested yet.

A single blended difference tells you the account lost. It never tells you where.

Match the two records on the same axes. Count the signals the test produced while you were at the desk. Then count the ones you took. A gap there is drift, and the missing trades are usually the uncomfortable ones, which are often the good ones.

Then split both records by hour. Say the tested edge lived in the first thirty minutes. Check whether your fills in that window are two ticks worse than the rest of the day. If they are, the problem is cost, with a time of day attached.

One more comparison worth running: average winner and average loser in R, live against test. Same win rate with a smaller average winner puts the blame on your exits.

Ten steps. The first six happen before you place a single live order.