Thibauld Charles Ghislain Robin / 프로필

Lately, I’ve been working on someone else's project with a very specific, and difficult, challenge: How can a model learn to trade with very few signals (say, 2 to 3 per month) and almost no data?

At first, it sounds simple. Less data, fewer variables, fewer decisions to make. But the opposite is true, the less you give to a model, the harder it becomes to extract anything meaningful.
When information is sparse, everything matters, the architecture, the way you define a “signal,” the filtering, the preprocessing. There’s no room for noise.
That’s where it gets complex. Not in quantity, but in depth. And the real enemy here is overfitting.

With limited information, and even more generally, a model can easily fall into one of two traps:
-It becomes too conservative, acting only when conditions are perfect. Safe, but useless.
-It starts memorizing the past, fitting perfectly on seen data, and collapsing the moment anything changes.
This project has been a reminder, building a model isn’t the hard part. Making it generalize is.

How to Detect Overfitting?

There’s a simple, brutal test: run it on a different asset.
If the model holds up, even slightly, it means it has found something real. Not just some random historical patterns. If it falls apart immediately, it never learned anything useful. It just repeated history.
That should be the first filter for any strategy, the ones you see everywhere: can it survive outside its comfort zone?

To fight overfitting, we use a few techniques:
-Data augmentation: shifts, scaling, noise injection, time warping… ways to show the model new versions of the same reality
-Synthetic data: simulated market data that retain structure while testing the model’s ability to generalize
-Regularization: constraints to avoid excessive complexity, forcing focus on core dynamics
-Cross-asset validation: (for me, the most critical method in trading) training on one or multiple assets, testing on others
But here’s the truth, even with all that, sometimes it’s not enough.
There are problems you can’t solve with more data or smarter code. Sometimes, what makes the difference is judgment, knowing when something looks too good to be true, when a backtest is lying, when to discard a “perfect” result. That’s where experience matters.

That’s where real trading begins and the theory ends.

Despite all the complexity, one truth remains: put it in a new environment. If it still breathes, it might be real.
That’s exactly why I trust the free strategy I released recently. It was trained on a single pair, live-tested, and it holds up reasonably well on some others. Not flawlessly, but consistently. That’s not luck. That’s a signal. That’s an edge.

And let’s not forget, when we talk about hourly data going back to the year 2000, we’re dealing with around 156,000 bars. That’s little. In machine learning terms, it’s barely a drop. And it gets worse, those 156,000 bars assume the market has never changed. No structural breaks. No volatility shifts. No regime changes.

So the idea that a model could “learn the market” from this alone is optimistic at best, dangerous at worst.
Which is why generalization, not memorization, must be the core goal and there's only one test to know if it's able to.

As for the freelance project, I haven’t cracked it yet. And it’ll give me hard times.
But the approach feels right. Sometimes, the hardest constraints lead to the most interesting breakthroughs.