Contents
What an edge actually is
An edge is a measurable, repeatable asymmetry between your decision process and the outcome distribution of the market. It is not a high win rate and it is not a secret indicator. A 35 percent win rate with a 3:1 reward-to-risk ratio is a strong edge. A 70 percent win rate with a 0.4:1 ratio is a losing one.
The word that matters is measurable. If you cannot state your edge as a rule that produces a signal, and you cannot count the signals over a large enough sample, then you do not have an edge. You have a feeling that has worked recently.
This is where most self-taught traders get stuck, because the search for an edge feels like the search for a secret and it is actually the search for a specification. The specification is a set of conditions under which you take a specific action with a specific invalidation.
Start from a hypothesis about behaviour, not from an indicator
Indicators are transformations of price, so any strategy built purely on them is a strategy about price, and price-only strategies on a single instrument are heavily arbitraged. Behavioural hypotheses are different. They are claims about what market participants do in specific conditions.
A good hypothesis for gold might be that aggressive selling into a level that has held twice tends to be absorbed. That is a claim about behaviour, it is falsifiable, and it can be tested. A hypothesis like 'RSI divergences work better in gold' is a claim about an indicator, and it is much weaker.
The best behavioural hypotheses come from watching the tape long enough to notice something that repeats. Our tape reading guide covers how to develop that observation habit, and free order flow tooling covers the software side.
The specification, written down
Write the hypothesis as a checklist with no ambiguity. What is the setup condition, expressed in observable terms? What is the entry trigger, at what time and price? What is the invalidation, and where exactly? What is the target, and how is it decided? What is the maximum risk per trade? What conditions void the whole setup before entry?
If any line of that checklist requires a judgement call, it is not yet specified. 'Enter when buyers look strong' is not a specification. 'Enter when the 5-minute candle closes above the prior swing high and the delta on that candle is above its 20-bar average' is a specification.
The value of this exercise is not the document. It is that the exercise exposes how much of your edge is actually in your head and unspoken. Most traders discover that their supposed edge is two rules and a lot of improvisation.
Testing it without lying to yourself
The purpose of testing is not to find a profitable configuration. It is to find out whether your hypothesis survives contact with data that was not used to create it. A test that confirms what you hoped is not a test.
Split your data chronologically, not randomly. Random splits leak future information into the training period through autocorrelation, which is one of the most common ways a backtest lies. Use an in-sample period to refine and an out-of-sample period that you do not look at until the rules are frozen. Our guide to out-of-sample testing covers the mechanics.
Then walk forward. Rolling the window forward with expanding training data and a fresh test period is the closest you can get to live behaviour without losing money discovering mistakes. Walk-forward validation sets out the procedure.
Finally, apply real costs. On gold your spread and commission are a large fraction of a typical scalp edge, and a strategy that survives gross often dies net. We cover this in spread considerations for gold and risk-reward analysis.
Sizing: the part that decides whether the edge survives
An edge is not an edge until it survives the sizing process. A strategy with a genuine 40 percent win rate and 2:1 reward-to-risk has an expectancy of 0.4 units, but if you risk 5 percent of the account per trade, a run of six losses takes you down 26 percent and a run of eight ends the account. The edge is real and you are still dead.
Size on the risk you can survive, not the risk you can afford to lose. A rule like risking 0.5 to 1 percent per trade is not a superstition; it is a direct consequence of the longest losing streak your edge should be expected to produce. The Kelly criterion gives a mathematical version of the same idea, and we work through it in the Kelly criterion for gold.
Also size down when your edge is out of sample and size up only when you have evidence. The most common ruinous pattern is a trader who starts at half a percent, wins for a month, doubles to one percent, wins for a month, doubles again, and then hits a regime change with a size the data never supported.
Knowing when the edge is gone
Every edge decays. The market adapts, the participants change, the free lunch closes. The trader who does not have a process for detecting decay is the trader who will keep a broken system alive for years because it has a good track record in the back half of the sample.
Keep a rolling record of live performance by the same specification. If the live expectancy drops below roughly half of what your out-of-sample test suggested, and it stays there for a meaningful number of trades, that is a signal to re-examine the hypothesis, not to add size.
There is a version of this that matters for the whole market, and not just your system. A sudden structural change, like a broker switching to a different liquidity source or a shift in central bank policy, changes the behaviour your edge was based on. Regime detection is how you catch that before it costs you the account.