LessonThis is a lesson restated in new words. Any figure in it is an example from the source, not this desk's record.
A backtest fails when it is a search for examples that already went your way. Scrolling backward until a favorite candle pattern appears, then treating the next similar picture as a plan, is confirmation bias. A usable test starts from a fixed rule, then asks whether small changes in that rule still leave something worth trading.
Why it matters
Retail trading content often answers “what works?” with “whatever makes sense to you.” That sentence feels personal and sounds hard to argue with. It also refuses to say how a rule behaves when the future is uncertain, how large a loss is allowed to get, or how the rule was checked before money was at risk.
A story that feels logical is not an edge. Betting black because black “looks due” is a feeling about a roulette wheel, and the wheel does not pay you for feelings. Markets work the same way. A reason you can narrate on a chart is not the same thing as a positive expected value, the probability-weighted average outcome of the bet.
Public personalities can be excellent showmen and weak teachers of this distinction. Wealth displayed on camera can swing, in the content itself, from one inflated label to another. The posts that actually discuss trades often draw far less attention than lifestyle clips. None of that is a probability model.
The idea in plain words
Confirmation bias means you notice the cases that support the story you already like. In a chart, that looks like rewinding until an “order block,” a gap, or a break in structure lines up with a move you wish you had taken. The past trade is selected because it worked. The future is not obligated to repeat the selection.
A proper historical test does the opposite. You write the rule before you look at the next result. You choose a model and a set of parameters, the knobs inside the rule. Then you move those knobs. You want to see whether a tiny change makes the history collapse, whether the result only softens a little, and whether there is any tradeoff left that you would actually accept.
Fitting noise is what happens when the rule is rebuilt to match accidents in one sample. It is easy to do by accident if the test is “find one or two nice trades from last week and memorize the picture.” It is much harder to do if the rule is frozen and then stressed.
The sports version makes the emptiness obvious. Saying the secret of football is that blockers block and the ball carrier either runs or throws is true and useless. Pointing at one play where running scored does not tell you the next play. Candle folklore has the same shape: the vocabulary sounds specific, and the decision rule is still “it looked right.”
How to use it
Write the entry, the exit, and the size in sentences that another person could follow without watching you click. Include the market, the timeframe, and what would cancel the trade. Run that exact rule on history you did not use to invent it.
Then perturb it. Shift a threshold, a lookback window, or a stop distance by a small amount. If only one narrow setting looks good and its neighbors fall apart, you have fitted a story, not found a stable rule. If a neighborhood of settings is merely less good, you at least have a tradeoff to discuss.
What usually goes wrong
The common failure is a loop: watch a week, highlight the trades you wish you had taken, name the building blocks, and repeat until the names feel like skill. That loop rewards pattern recognition after the outcome is known. It does not estimate how often the same picture fails.
Another failure is blame-shifting. If the rule is vague, any loss can be filed under psychology, “you saw the gap wrong,” or “trading is personal.” Education that cannot be wrong is not education. A test that can be wrong is the point.
A third failure is treating a show as a syllabus. Attention is not a research method. A method that will not get a huge audience can still be the one that tells you whether a rule was noise.
Questions people ask
Is a strategy valid if it feels logical?
No. Logic is how you state a hypothesis. Validity is whether the stated rule has a positive expected value after you stop picking the examples.
What is the difference between review and a backtest?
Review asks what you notice. A backtest applies a frozen rule, then changes its parameters to see whether the result was a single lucky setting.
Why do past “perfect” setups mislead?
They were chosen because they worked. That selection is confirmation bias, and the next chart is not required to match the gallery.