How to build a trading strategy, and how to break it first

Most advice on building a trading strategy is about where ideas come from. Ideas are the cheap part. Anyone can produce a dozen plausible rules in an afternoon, and a backtest will happily tell you several of them worked. The expensive part is elimination, and it has an order to it. Each stage below can invalidate everything after it, so running them out of sequence wastes the work. What follows is the order we use, with the measurement behind each step, because every one of these is a way we have watched a backtest tell us something untrue.

1. Write the abandonment condition first

A strategy that has no stated failure condition cannot fail. It can only be adjusted. Decide in advance what you are testing, over what period, with what costs, and what result would make you throw it away. Then keep that document and do not edit it after seeing results. Everything in this guide is a way of meeting that document honestly; none of it helps if the target moves.

2. Your universe is wrong before a single trade is simulated

Universe construction is where most of the damage happens, and it happens silently because a broken universe still produces a clean-looking equity curve. Three failures we have measured, each of which survives casual inspection: The survivorship case is the one people nod at and then get wrong anyway. It is not a random 1.75%: the names that disappeared traded at a median of $17.11 against $35.99 for the survivors, so the missing days are concentrated in the cheaper half of the market. Details in survivorship bias, reverse splits in gainers lists and what a public screener misses. The fix is mechanical: build the universe from the data of the day being tested, never from a current list. If you cannot, say so as a limitation rather than treating the result as complete.

3. The prices in your file were not tradable

A bar file gives you four numbers per period and no indication of whether any of them were available to you. Any rule that reads a price and assumes it was actionable inherits all three. The measurements are in the last price is not a price you can trade at, order book depth and dead minutes.

4. Fills are where the fiction lives

A fill assumption is invisible in the output. Nothing in an equity curve tells you whether the trades behind it could have happened, which is why this survives longer than any other error in the list. The test is to take the filled trades and ask, one at a time, what would have had to be true at that moment for the fill to occur, then check whether it was. Most backtests never ask, because the simulator answered the question already by assuming it. Two rules remove most of it. First, a limit order at the best price is not a fill; something has to trade through your price, and you are behind whatever was already resting there. The reasons it does not fill are enumerated in why your limit order did not fill. Second, a market order does not execute at the quote you saw: a $5,000 market buy in a $1-5 name costs 88bp one way against 45bp for a $1,000 order, and a $25,000 order could not be filled inside ten visible levels 28.7% of the time (market order vs limit order). If your backtest fills at the touch, it is not modelling execution. It is modelling a market that was waiting for you.

5. Cost models are wrong in both directions

Everyone knows to include costs. Fewer people check whether the cost model is right, and the error runs both ways. A flat percentage is wrong because commissions do not scale with size: the same $1,000 order costs 0.35bp on a commission-free broker and 24.04bp on a per-order schedule, a ratio of 68.7x. Which broker you assume decides the answer before the rules do. The error also runs the other way, and that direction is the dangerous one because it produces rejections rather than acceptances. Regulatory fees change on their own schedule, sometimes to zero for months at a time, and a model carrying a stale rate will charge trades a bill that was not payable. An idea killed by a cost assumption is killed quietly, with no result to inspect. Both directions are worked through in how to model trading fees and what commissions actually cost. Before comparing anything to anything, know the toll. A $5,000 market buy in a $5 to $20 stock pays roughly 0.75% round trip against the mid before commissions (the toll every trade has to beat). An idea that produces less than that is not a marginal strategy; it is not a strategy.

6. Test the exit as hard as the entry

Entries get the attention because they are where the idea lives. Exits decide the result and are usually specified in a sentence. The exit is also where the market is least accommodating. On a fast drop the spread widens and the book thins at the same time, which is precisely when a stop fires (stop-loss vs stop-limit). Size the position for the book you will leave through, not the one you entered through (how to size a day trade). A related trap is holding periods. If your universe turns over during the session, a rule that assumes the name stays interesting is not being tested on the same population it selected: of names on a 30-slot list at 09:35, 57% were still there at 10:00 and 33% by the close (watchlist turnover).

7. One held-back period is not out-of-sample

The usual practice, fit on one period and confirm on a later one, is necessary and not sufficient. If you have looked at the second period even once and then changed anything, it is no longer out-of-sample; it is a slower version of fitting. Two things help. Freeze the rules in writing before the second period is opened, including the exact thresholds, and hold back a third period that nobody touches until the end. And distrust results that rest on a handful of trades. A profit factor computed over single-digit trade counts swings between extremes on the same day with the same rules, which is information about the sample and not about the strategy.

8. What a pass actually looks like

Not a high return. A result that survives every step above while you were actively trying to break it, and that still works when you assume the least convenient fill, the highest plausible cost, and the universe as it existed on the day rather than today. Most ideas do not get there. That is the normal outcome and it is not wasted effort, because the elimination teaches you something about the market even when it teaches you nothing about the idea. Often the useful residue is a piece of folklore that turns out to be false. The claim that gaps always fill does not survive contact with 144,286 of them, and knowing which part of it is false is worth more than a rule built on the part that is true. What this guide assumes That you already have a backtest that runs, and that your question is whether to believe it. If you are earlier than that, the gaps between a backtest and live trading are enumerated with costs attached in why your backtest does not match live trading, which is the shorter version of this argument. None of the above will tell you whether a strategy will make money. It tells you whether the number in front of you is measuring what you think it is measuring, which is a different question and the only one a backtest can answer.