It is surprising how seemingly very small adjustments can make a big difference. I only recently figured out how to export data from my broker and import into EA Studio. I can't believe how just changing the datasets really impacts some strategies.

Also the Walk Forward is great. I have a lot of strategies from Generator that pass through all my criteria but sometimes they just look like crap on the WFA. So I use that Walk Forward as a filter for all top strategies now.

I've also learned a lot about trading and understanding statistics by studying the reports and then asking AI chatbots questions like:
In EA Studio, what does the column SQN represent? What should I look for in the Win/Loss category? What about Return/DD?

And how many of these robots are actually performing live, Vincenzo?
Because I’ve seen your live accounts during the Zoom call you even admitted that almost all of your profitable accounts are grid systems.
No need to defend it, amigo.
You know it, and I know it that’s enough.”

Hi Blaiserboy,

WFA can really open your eyes. It’s often the moment you realize which strategies are truly robust and which ones just look good on paper. Or, the moment you get unexpected surprises.

But, I also run 20+ demo accounts (the oldest over 500 days, three just opened this week) with 1,000+ EAs running at any time to validate workflow, setup, and criteria changes. The demo stage is another gatekeeper — and then, of course, the live environment puts everything under discussion again.

I treat this as a pipeline with gatekeepers:

Gate 1: Generation + Robustness tests

Gate 2: Walk-Forward Analysis in EA Studio

Gate 3: Demo (≥180 days, 50+ trades)

Gate 4: Live (≥30 trades)

My view is, if you create rubbish at Gate 1, your success rate at Gate 4 becomes unmanageable. With only 20–30 % EAs success rate, it’s almost impossible to maintain a profitable portfolio long-term.

But if you filter properly along the way, you only need around a 60–70 % success rate. In a 10-EA portfolio, that’s usually enough for 6–7 to perform well — and that’s what makes the difference.

The road is long and not easy.
I’m sharing what I do not to teach, but to be constructively challenged and learn from others’ experience too.

I’ve been trading for over 10 years, almost 5 of them working with bought, selected, and “domesticated” EAs. What I’ve learned (also thanks to others sharing) is that data sustain results.
The time you invest in studying and testing is much larger than the time needed to manage portfolios.

If we all share a bit of experience — supported by data — we could all learn faster and make this journey more sustainable. Or we could even work together to validate ideas.

Thanks
Vincenzo

And to prove that you’re not an expert and that you actually proved my point of what I said in my topic, that short data isn’t robust I asked ChatGPT about the exact same experiment as if it were mine.
I did that deliberately, presenting it as my own experiment, to get the most neutral and objective answer possible.
I already knew the outcome when I saw your post, but since you keep saying that ChatGPT “opens a whole new world” for you, I asked it myself.

Look and learn, Vincenzo the expert.

NO SUGARCOATHING FULL TRUTH I WANT THE HARD TRUTH IS THIS THE ROBUST WAY ON DATA 2016 TO 2023 Experiment #1 — Can the filters you use to select EAs impact Walk-Forward Analysis (WFA) success?
I took the same collection of 233 EAs, the same dataset, and the same tool.
This collection was built early in my EA Studio journey — nothing special at all.

Then I ran it over and over, each time applying a different combination of filters before validation.
After filtering, the remaining strategies were tested on unseen data (WFA).
In total, I ran 91 combinations, each representing one complete backtest + WFA cycle.

The results: (attachment)

Some combinations achieved 100 % WFA success, others dropped below 20 % — all with identical data and strategies (Example: 50 profitable in WFA out of 50 that passed backtest = 100 %.
20 profitable in WFA out of 100 that passed = 20 %.).
ChatGPT zei:

Alright bro, no sugarcoating — let’s get surgical with this.

First: What you actually tested

You didn’t test robustness of strategies.
You tested robustness of your filtering logic.
Meaning: you proved that filter choice alone can completely distort your perception of robustness, even when the underlying data and EAs are identical. That’s a powerful insight — but it also exposes a critical weakness.

The hard truth

This experiment does not prove robustness.
It actually proves the opposite: that your results can be manipulated — consciously or not — through selective filtering. Here’s why:

You used one fixed EA dataset (2016–2023).

That’s only one market era.
You have no independent data segment from before 2016 (e.g., 2008–2015) to verify true generalization.
Without pre-2016 exposure, you can’t see how those systems handle different volatility regimes, spreads, or liquidity conditions.

Walk-Forward (WFA) on 2016–2023 is still “in-sample-ish” because all EAs were born from that structure.

Your sample (233 EAs) is too narrow and too correlated.

If most were generated from the same timeframe, symbol, or parameter pool, they share structural biases.
You’re not testing 233 independent edges — you’re testing variations of the same bias.
So your 100 % WFA success pockets can easily be statistical mirages.

91 filter combinations = 91 different curve fits.

Each filter set changes the data-mining bias.
A few will accidentally align with the validation set — giving fake “100 % WFA success”.
That doesn’t mean those filters are robust; it just means they fit the validation noise better in that run.

WFA success ≠ future performance.

WFA can validate the process, not guarantee robustness.
Even perfect WFA results collapse live if your generation set lacked diverse conditions.
You can get 100 % WFA pass and still fail OOS (out-of-sample) if your entire data span is homogeneous.

The deeper reality

Robustness only exists when your strategy creation → validation → live pipeline consistently survives multiple uncorrelated stress environments.

Here’s the real robust structure (truth-level setup):

Phase    Data    Purpose
Build    2008–2019    Learn every environment possible (crash, recovery, QE, COVID)
Validate (WFA)    2020–2023    Forward stability check
Test (OOS)    2024–2025    Proof of generalization
Live monitor    current    Reality check

You only covered the middle phase — not the full robustness chain.

Brutal verdict

No, this is not the robust way.

It’s a methodology stress test, not a robustness test.
You’ve demonstrated how fragile robustness appears when filter design is arbitrary.
True robustness means proving consistency across:

Different data spans

Different market conditions

Different generation sessions

Different brokers & spreads

Different filter sets (and still similar outcomes)

Until your process can do that — you’re seeing artifacts, not truth.

What is the robust way then

Generate new sets for each data regime (e.g. 2008–2015, 2016–2020, 2021–2025).
Cross-validate them. A robust logic will keep showing up across all eras.

Randomize and stress test each with Monte Carlo on:

Data shifts

Spread variations

Execution noise

Filter only on universal properties (PF, DD, SQN, #trades, smooth equity).
Avoid narrow or aesthetic filters — they create false robustness.

Validate on unseen broker data (not just unseen time).
True robustness = cross-broker stability.

TL;DR

2016–2023 WFA alone = not robust.

91 filter combos = proof of bias, not strength.

100 % success = likely curve-fit filter alignment.

To prove robustness, your system must survive different market DNA, not different filters.

Truth: you just proved how easy it is to fool yourself.
And that’s the most valuable lesson you could’ve extracted.

Would you like me to show you the exact pipeline structure (data splits + validation layers + stress tests) that would make this experiment truly robust?

Reminder to All Participants

In order to keep our discussion constructive and focused on ideas rather than individuals, please avoid referencing personal qualifications, credentials, or experiences in your responses.

Hi Blaiserboy,

I’m not sure if I got your question exactly right, but let me explain how I use WFA — maybe this helps.

I actually do different things depending on the phase:

Reactor (generation phase) →
When I generate strategies, I apply a very light optimization (only TP/SL, no indicator parameters) with 20% OOS.
I also always leave out the last 6–12 months completely.

Validation phase →
When I validate existing collections or single strategies, I backtest from 2016–2024, then filter to pick the top performers and run a full 2025 validation (WFA). Here there’s no need to split into segments, since that data was never used before.

Reactor (new setup) →
I recently launched a new creation batch using Walk-Forward Optimization (WFO and no additional optimization) with 10 segments and 40% OOS. This one will run until the end of the week — then we’ll see the results.

Did that answer your question?

That gives me an idea or two

Thanks very much