There is a particular kind of effort in technical analysis that masquerades as rigour while quietly undermining it: the relentless adjustment of indicator settings to maximise performance against historical data. The activity feels responsible. One tests a moving average of one length, then another, then a momentum threshold of one value, then a neighbouring one, recording how each variation would have performed and converging steadily on the combination that produced the best historical result. The trouble is that this process, carried far enough, does not discover a better strategy. It discovers the settings that happen to match the specific accidents of the past, and the past, however carefully studied, never repeats itself precisely.
This is the phenomenon known as over-optimisation, or curve fitting, and its danger lies in how convincing the results appear. A strategy tuned across hundreds of variations to fit a particular stretch of history will display a beautiful equity curve over that stretch, because it has been shaped to do exactly that. The settings that produced the curve, however, encode not the durable structure of the market but its random noise — the particular sequence of moves that will never occur again in the same order. When the strategy meets new data, the noise it was fitted to is gone, replaced by different noise, and the performance collapses. The investor is left holding a tool optimised for a world that no longer exists.
The deeper problem is that over-optimisation exploits the freedom inherent in having many parameters to adjust. The more variables a strategy contains and the more finely each can be tuned, the more certain it becomes that some combination will fit the past well by pure chance. A system with enough adjustable settings can always be made to look excellent in hindsight, regardless of whether it captures anything real, in the same way that a sufficiently flexible line can be drawn through any scatter of points. The impressive historical result, in such cases, is not evidence that the strategy works but evidence that it has been given enough freedom to memorise the data, which is the opposite of understanding it.
The defence against this trap is a deliberate preference for robustness over perfection. A strategy worth trusting is one that performs reasonably across a wide range of settings rather than spectacularly at a single precise combination, because such broad competence suggests it has captured a genuine tendency rather than a historical accident. If shifting a parameter slightly causes performance to crumble, the strategy was balanced on a knife edge of coincidence and will not survive contact with new data. If performance degrades gracefully as settings change, the strategy rests on something more durable. The disciplined investor therefore seeks the configuration that is merely good and stable, treating any setting that appears extraordinary with suspicion rather than satisfaction.
What over-optimisation ultimately reveals is a confusion between explaining the past and anticipating the future, two activities that pull in opposite directions. A strategy can be made to explain the past with arbitrary precision by adding parameters and tuning them, but every increment of fit to history is purchased at the cost of fit to whatever comes next. The goal of analysis was never to construct the tool that would have performed best over a chosen stretch of data; that tool is always available and always worthless. The goal is to find an approach simple and stable enough to keep working when the future arrives wearing a face the past never showed, and that goal is served by restraint, not by the endless refinement that so often passes for diligence.