Strategy Development Is Easy Now. Validation Is Not.
Roll back just one or two years and strategy development itself was a problem.
You had to sit down, code it, craft it, debug it, design the whole thing properly.
That problem is basically gone now.
Today I can build pretty much any reasonably defined strategy in under an hour. A lot of them take 15 minutes.
Funny enough, it may now take longer to describe the strategy properly than to actually code it.
And that is where requirements engineering suddenly becomes gold.
If you can clearly explain what you want, AI can usually code it very quickly. If you cannot explain it properly, AI will happily start making decisions for you. Sometimes those decisions are sensible. Sometimes you don't even notice it made them.
But coding was never really the hardest part.
The real question is:
What happens after the strategy exists?
My current process looks something like this:
Idea → Code → Verification → Edge Survey → WFO → Sim → Live → Monitor → Rotate
AI can reduce the workload in almost every one of these stages.
What it cannot do is remove the stages.
1. Idea
Everything starts with an actual trading idea.
Not:
"Let's optimize 17 indicators and see what makes money."
An idea should be something you can explain in a couple of sentences.
For example:
Markets that expand strongly after a compressed overnight range may continue in the same direction after the open.
That is something we can test.
At this point I don't care about finding the perfect parameters. I care about whether the basic concept has any edge at all.
2. Code
This is the easy part now.
Once the rules are clear, AI can normally produce the first implementation extremely quickly.
Entries, exits, sessions, filters, position sizing, whatever.
A few years ago this stage could consume most of the project.
Now it is almost infrastructure.
But fast code creates another problem: you can produce bad ideas much faster too.
So the fact that something can be coded in 15 minutes does not mean it deserves another three days of research.
3. Verification
Before researching performance, I want to know that the strategy is actually doing what I think it is doing.
This sounds obvious, but it is probably one of the most ignored steps.
Check trades manually.
Check timestamps.
Check entries.
Check exits.
Check session boundaries.
Check stop and target behaviour.
If the strategy specification says one thing and the backtest engine is actually doing something slightly different, every result after this point is garbage.
AI helps massively here because you can ask it to inspect logs, trade lists and code together.
But you still need to verify the behaviour.
4. Edge Survey
This is probably the most important stage.
I call it an Edge Survey because I am not trying to optimize the strategy yet.
I am trying to answer one question:
Does this concept appear to contain an edge?
This is also where strategy design becomes difficult.
The strategy should not contain an enormous number of combinations and filters.
Otherwise you are not testing one idea anymore.
You are effectively testing hundreds or thousands of slightly different strategies at the same time.
Then something inevitably looks good.
And this is where research becomes dangerous.
You tweak a parameter.
Run it again.
Change another filter.
Run it again.
Add another condition.
Run it again.
Twenty runs later you barely remember what the original idea was.
Come back the next morning and sometimes it feels like starting the research again from scratch.
So the boundary has to be clear:
This is the concept I am testing.
If it does not show enough evidence of edge at this stage, throw it away and move on.
That is much harder psychologically than coding another filter.
5. WFO
If the basic idea survives the Edge Survey, then I can start looking at parameter stability.
Walk Forward Optimization is not there to magically find the best settings.
For me it is much more about asking:
Does this thing still work when the market changes and the parameters are periodically reselected?
I want reasonable neighbouring parameter values to behave reasonably.
I don't want one beautiful parameter combination surrounded by disaster.
AI can help compare windows, parameter sets and degradation patterns, but it cannot change the basic principle:
A robust strategy should not depend on discovering one magic number.
6. Sim
After research comes simulation.
Now I want to see the strategy operating in the actual environment.
Live data.
Real session transitions.
Actual platform behaviour.
Orders being submitted and cancelled.
Data interruptions.
Restarts.
All the boring things that a backtest does not properly reproduce.
This is also where you start finding differences between a trading model and an actual trading system.
Those are two different things.
7. Live
Only after that do I want real money involved.
And preferably not much of it.
Going live should not suddenly become another experiment.
By this point I should already understand roughly what the strategy does, where it struggles, what kind of drawdowns are normal and what kind of behaviour would make me concerned.
Live trading is mainly confirmation that reality still resembles the research.
8. Monitor
This is another area where I think AI becomes extremely useful.
Strategies should not just sit there running forever because their backtest looked good six months ago.
You can continuously compare live behaviour against research expectations.
Trade distribution.
Drawdown.
Win rate.
Average trade.
Market regime.
Execution quality.
Parameter behaviour.
Instead of manually opening five reports and trying to remember what happened last month, AI can maintain the context and tell you what actually changed.
That, to me, is much more interesting than asking AI to invent another strategy.
9. Rotate
Eventually strategies change.
Some stop working.
Some become weaker.
Some only work in certain environments.
Some get replaced by better versions.
So rotation is part of the system too.
The goal is not to find one magical strategy and trade it forever.
The goal is to maintain a collection of researched edges and understand which ones deserve capital right now.
And this is where I think AI actually changes systematic trading.
Not because it can write a strategy in 15 minutes.
That part is already becoming boring.
The interesting part is using AI to maintain research memory.
It can read a 300-page quant book and explain the model back to you in baby language.
If you have a systematic mindset and school-level mathematics, a huge amount of publicly available quantitative research suddenly becomes usable.
But that still doesn't solve validation.
The advantage comes when AI starts helping us remember every experiment, every parameter set, every failed idea, every walk-forward run and every live deviation.
Because the biggest problem in strategy research is often not computing the result.
It is keeping track of why you ran the test in the first place, what changed, and what the result actually means.
Strategy generation is becoming cheap.
Good validation is not.