Do Better
Farming Trade Data To Feed AI For Better Strategy Performance
There are no guarantees that any performance you see here will continue in the future. I recommend using ATS strategies in simulated trading until you/we find the holy grail of trade strategy. This is strictly for learning purposes.
I’m not just developing strategies—I’m on a quest for the holy grail of automated trading. Questions? Check the FAQs first, then feel free to reach out directly: AutomatedTradingStrategies@protonmail.com.
“AI will create a new science.”
Sam Altman, CEO of OpenAI
The AI Arms Race
It’s official, the AI race is on. Sam Altman, the CEO of OpenAI is teasing Chat GPT 5’s release this summer and I’m one of the idiots that’ll be paying whatever it costs to access. I’m currently spending ~$1K a month on AI related models/products and I use them for everything. In addition to building an RL Strategy Prediction Model for ATS, I also use it for bird identification, cooking, bushcraft, ancient text translations and legal analysis. I’m even using it as a way to store, track and analyze bloodwork for my family (including pets) —longevity escape velocity here we come.
I use multiple AI models to prevent hallucinations and echo chamber errors, but I’m also just curious about how each model differs, especially with regard to tool-set and automated strategy development.
Not surprisingly, the competition is heating up. From infrastructure build out to AI talent, all that dry powder out there is finding its way to AI. According to Altman, Mark Zuckerberg is offering members of OpenAI’s research team $100M sign-on bonuses. And last week, Meta officially purchased a 49% stake in ScaleAI for a whopping $15 billion. Why?
The “D” word.
It’s all about DATA—labeled data in particular.
This is no surprise because I’m finding the same needs in my own research. Market data is great for initial strategy development, but trade data, actual trade data, can help turn an unprofitable strategy into a winner. It’s not just about finding the holy grail, it’s about finding the strategies that are most likely to be profitable on any given day.
The more data I collect, the better the analysis gets, which is why I’ve added two forward test Incubators in Q2, and I’ll be adding a third in Q3 and Q4. I’m currently running over 200 strategy variations.
And don’t forget about derivative data points. My new favorite metric is ‘% of trades greater than average MAE’. I’ve got one strategy in the Incubator with 7%. That means that only 7% of trades had an MAE that was greater than the avg MAE, which in this case was $400. This is risk management gold. You’ll be hearing a lot about this metric in the future, especially with regard to stop placement.
And yet today, as I roll out Strategy 94v3, I know that only 50% of the strategies have been forward tested. In Q2, I started forward testing strategies 1 through 50. In Q3, I’ll add strategies 51 through 100. The data file is growing. The AI models are getting better. The analysis is getting better. The testing is better. The output is better. And, my overall performance is getting better. Today, I can ask questions like, “which strategies are 100% profitable on Wednesdays” or “which strategies have a 100% win rate between 10 am and 11:30 am, or one of my favorites, “which strategies have an 80%+ chance of mean reverting after a negative day”?
This is why the first thing I tell everyone to do is: start a forward test. I know it’s not what folks want to hear. I suppose you could use backtest data, but garbage in garbage out. If your prediction fails, you won’t know where to concentrate your effort: did it happen because of poor historical data, a bad simulator, or bad prediction logic? There is nothing better than actual trade data, especially for high frequency or scalping strategies like Strategy 94v3.
Scenario’s 5 & 6: The Evolution Continues
Trading strategy development has traditionally relied on intuition, backtesting, and incremental adjustments. ML has given us the ability to study a different approach: using comprehensive trade data analysis to completely rebuild a strategy from its core principles. Strategy 94 was the second iteration of this attempt, but with every attempt, we get more data and ML loves more data.
The transformation of Strategy 94 into Strategy 94v3 represents a continuation of a data-first development process that lets performance metrics—not assumptions based on backtests—drive design decisions.
In total, there were 23 variations of 94v2 running in various forward tests in Q2. Each scenario varies with the use of different filters. I’m going to share their performance with you today. We’ll look at:
which filters performed the best and why.
which bar calculation (bar close or each tick) performed the best and why.
which time series performed the best and at what times.
We’re also going to get into trade duration statistics because 75% of Strategy 94v2’s trades hit the target profit in under 5 minutes. Strategy 94v3 incorporates this stat into the trade logic. Strategy 94v3 is available for download at the bottom of this post.
Understanding Strategy 94v3 Filters & Features
Based on the analysis of 4,523 trades across 23 Strategy 94 variations, I've identified critical performance drivers and optimal configuration settings based on actual trade data. Here’s an overview of performance metrics from all variations.