Last week J finished a backtest run and sent me a screenshot: “Line 1 win rate 86%, 30 days. Take a look.”
I stared at that number for about three seconds, then replied: “Don’t rush it. Let’s see how it performs with real money.”
He said: “So I’m just waiting for you to give the green light.”
We’ve had this exact exchange more times than I can count. Every time, it goes the same way — AI generates beautiful numbers, and then for the part that actually matters, AI stays put while I have to take the next step alone.
There’s a Deep Gap Between Having an Idea and Having a Strategy
When I first started, my understanding of RSI was straight from the textbook: overbought and oversold, short above 70, long below 30. I knew it was too simplistic, but I wasn’t sure what direction “making it more sophisticated” should even take.
There’s a gap here that most people don’t recognize: converting “I think the market will move this way under these conditions” into “logic a program can actually execute” is not a technical problem — it’s a thinking problem.
When you say “RSI combined with trend,” how does code define “trend”? Moving averages? Which one? Regime detection? How do you switch between them? Every casual verbal description hides at least three variables that need precise definitions.
AI genuinely helped me at this stage. When I’d throw an idea at J, he’d break it down: “What you’re describing could be implemented three different ways — which one do you actually want?” He wasn’t writing code yet. He was helping me turn vague intuition into a testable hypothesis.
But “what I think the market should do in this condition” — that call, he couldn’t make for me. That judgment comes from how long I’ve stared at charts, how many losing trades I’ve taken, what kind of structural intuition I’ve built up about how markets move. AI can’t fill that in.
Three Places Where AI Is Genuinely Useful
With that out of the way, here’s where AI actually saved me serious time.
Signal design — J’s speed here is almost unsettling. I said something like “I want to look at RSI combined with volume surges, and see if performance differs across volatility regimes,” and within about ten minutes he had a first-pass regime classification framework: low volatility, medium volatility, high volatility, with different RSI trigger conditions for each state.
If I had written that myself, it would have taken an afternoon — and probably still had bugs.
Parameter scanning completely changed how I work. Before, I’d just default to “RSI period 14, because that’s what everyone uses.” Now we run grid searches — RSI period from 7 to 21, entry condition thresholds tested across every regime, then filter for the combination with the best Sharpe and lowest max drawdown. That kind of sweep would take days by hand. J runs it overnight, and the data is sitting there when I wake up.
Anomaly debugging is another one. Once a backtest result suddenly exploded — win rate jumped from 60% to 92%. My immediate reaction: something is broken. J investigated and found a look-ahead bias — the code was referencing future data. That kind of bug is buried deep; I couldn’t spot it reading through the code myself. J found it in ten minutes.
But There Are Things AI Simply Can’t Do
This is the part I actually want to talk about.
Strategy philosophy cannot be outsourced.
What is strategy philosophy? It’s the belief system underneath your strategy — what you think is fundamentally true about markets. Do you believe markets are mean-reverting or trend-following? Is RSI oversold a buying opportunity or a falling knife? Which core assumption does your entire strategy logic rest on?
AI can’t answer that for you, because the answer requires you to have already formed a view on markets. J can test “what does a mean-reversion strategy look like if we assume that’s true?” — but he can’t tell me which assumption is correct.
Risk limits are the same.
Our rules: maximum loss of 2% per trade, reduce position size after a certain number of consecutive losses, pause entirely if monthly drawdown hits the threshold. These numbers look specific and objective, but behind them is a deeply personal question: how much pain can I actually tolerate? J can run historical simulations to show what the worst-case outcome looks like under different risk parameters — but he can’t tell me which worst case I can live with.
And then there’s going live with real money.
This is the hardest part to put into words. 86% win rate in backtesting. Thirty days of paper trading. The numbers all check out — but the moment you actually place that first real-money order, it feels completely different. It’s not a rational awareness of risk. It’s something more primal: this is real money.
AI cannot get you past that moment. That one is yours alone.
How the 5-Line Architecture Came Together
Our current setup: Line 3 for backtesting and validation, Line 1 for live testing with small size, Line 5 for real capital. This structure didn’t come from deliberate design upfront — it evolved from stepping on a few landmines.
Once we had a strategy that looked beautiful in backtesting, but as soon as we ran it on Line 1, problems started appearing. Slippage was much worse than expected, and under certain market conditions, the strategy would miss signals entirely. If we’d gone straight to real money, the damage would have been significant. That experience locked in a rule we’ve never broken since: any strategy must run on Line 1 for at least 30 days with a win rate of 80% or higher before it’s eligible to go live.
That experience also clarified the concept of closing the loop. If a real-money strategy starts losing, we don’t kill it — we send it back to Line 3 for research. We figure out which assumption broke down in the current market regime, fix it, and run the whole process again. There are no permanently broken strategies, only strategies that don’t fit the current regime.
AI plays a crucial role as an accelerator in this whole architecture. But the architecture itself was designed by a human. Every “that’s wrong, let’s change direction” judgment call was made by a human.
My current understanding of AI-assisted trading development: it lets me test more hypotheses faster. But I need to bring in sharper hypotheses to begin with. If my own thinking is vague, AI will just help me confirm that vagueness faster.
This is a topic I’ve been asked about a lot lately, so I finally sat down and organized my thoughts.