Overfitting: Poison, Illusion or Antidote?
In the world of quantitative trading overfitting is universally recognised as a threat. A pathology of the trading system. A mistake to be avoided. Overly beautiful equity curves, insignificant drawdowns, strategies that seem perfect… until they are confronted with reality.
But there is a more uncomfortable and more useful truth: overfitting cannot be eliminated.
The very moment one uses historical data to construct a strategy, some form of adaptation to the past is inevitable. Even the most sophisticated methods are not immune. And even a model conceived independently of the historian, but then tested and iterated, risks slipping into survivorship bias.
So the question changes: no longer how to avoid overfitting but how to use it to the benefit of the research process.
Statistical illusion, but also revelation
An over-optimised system is the product of over-optimisation. It is the result of hundreds of parametric combinations sewn perfectly on a past context. And like any poison, it may seem fascinating: outsized annualised performance, impeccable equity lines, rules that “always work”.
Yet in that very unreal perfection lie precious clues.
An overfitted model shows, paradoxically, how far a theoretical idea can go. It indicates the maximum (illusory) potential of a logic, but also its fragility. When analysed in depth, it can reveal:
- which variables determine the outcome;
- to which parameters it is hypersensitive;
- under which market conditions the “magic” is generated;
- and above all, which part of the performance is structural and which is pure illusion.
Like any poison, overfitting should not be ignored, but isolated and studied.
Controlled exposure
In biology, the antidote comes from exposure. It is not created by avoiding the poison but by injecting it in a controlled manner, observing the reaction and building defence on that response.
In quantitative trading, the process is no different.
Deliberately exposing oneself to overfitting, building models pushed to the limit, highly optimised is a useful and necessary practice. Not to use them directly but to study them as a controlled laboratory.
It is in this artificial space that extreme behaviour is observed:
- conditions are stressed;
- break thresholds are tested;
- sensitive zones are detected.
This type of work, if carried out methodically, activates the “immune system” of the strategy: it allows one to understand where to simplify, where to cut, where to reinforce. In essence, exposure becomes robustness training.
Accepting risk to build value
Only after crossing the unstable terrain of overfitting is it possible to build something truly robust. The antidote, in the quantitative process, is the result of an evolution, not a surrender.
A strategy that has faced and overcome overfitting knows its limits. It has memory of the contexts in which it fails, has learnt to distinguish signal from noise and has reduced its degree of dependence on critical parameters.
This results in models that:
- are generalisable to different instruments and markets;
- show consistent behaviour even outside the sample;
- maintain stable performance even in changed contexts;
- and above all, do not promise the impossible.
Robustness is not the absence of error: it is the ability to survive it.
Overfitting is an integral part of working in the market. It is not a mistake to be avoided, but a phase to be traversed with discipline and method. To attempt to eliminate it completely is to delude oneself. But to accept it as part of the process is to grow as researchers, as system builders, as conscious operators.
The real risk is not overdoing a model. The real risk is not knowing that you have done so, or worse, not learning anything from what has emerged.
In this sense, overfitting really does become a useful poison. Because if dosed and managed, it produces knowledge. And knowledge is the only real antidote.
