I built a trading system to understand markets. I understand markets. I still haven't traded.

It started because I wanted to understand how markets, trading and backtesting actually worked. Not use a platform, actually understand it. So I built a small engine from scratch one weekend. Entry on the next bar's open, no lookahead, costs deducted per trade. Clean, simple, done.

Except then I needed data.

So I pulled price history for the S&P 500. Then I thought earnings matter too, so I added that. Then insider trades. Then I built a news archive and tagged everything by sentiment and asset. Then a macro database with VIX, yield curve, credit spreads, fear and greed index, days to FOMC. Each piece felt necessary for the one before it. Some months later I had 20M+ rows across 587 symbols going back to 2018 and I was still adding things on weekends.

At some point I decided to train a model. I tested a lot of things that didn't work. Amazon's Chronos, a time series foundation model that gets a lot of hype, performed no better than a coin flip on strict non-overlapping walk-forward validation. Short horizons were pure noise. Predicting alpha vs SPY was inconsistent. What actually worked was a logistic regression predicting 20-day direction. AUC 0.54 out-of-sample. Not magic, but real and consistent from 2020 to today, which is more than I expected.

The main signal turned out to be macro. VIX, credit spreads, fed funds rate. Essentially mean reversion on fear. Gradient boosting matched the logistic regression, which told me the signal is linear and low-order. I kept the simpler model.

By then I had a backtest engine, a data pipeline, and a working model. The problem was that using it was painful. Defining a strategy meant filling out a long configuration form, and every time I wanted to try something new I'd spend more time on the config than on the idea itself. So I built a natural language layer: describe what you want, an agent parses it into structured conditions, and it runs. That turned out to be the part that made it feel like something other people might actually use.

I put up a demo at demo.botlance.ai. No registration, no credit card.

I've never invested a single euro with any of this. I don't know if that's ironic or just honest.

The thing that surprised me most was how little technical indicators mattered when training the model. I spent weeks adding RSI, MACD, Bollinger bands, convinced the edge was in the combinations. When the model finally trained, the features that actually moved the needle were VIX, credit spreads, and days to FOMC. The market is mostly driven by fear and liquidity, not patterns on a chart.

I also learned that a model with AUC 0.54 is not a bad model if you understand what you're measuring. Most people would throw that away. The right question is not whether it predicts correctly most of the time, it's whether the trades it selects perform better than the ones it rejects. They do, consistently, across six years of data.

I'm not sure yet whether this is a product or a very elaborate hobby. But if you've ever wanted to test a trading idea against real data without writing code, it's there. And if you've gone down a similar rabbit hole, building something for months before asking whether it should exist, I'd like to hear how that went for you.