Contents

I started this project meaning to build an automated trading system that would predict stock prices. Somewhere along the way, it turned into a tool for re-measuring how reliable the inputs behind my own decisions really are. Following the commit history in clusters of major changes, I want to look back at what I built and what I let go of.

First, a foundation for collecting data#

The first thing I wrote was plumbing, not analysis. Collecting stock prices and fundamentals, a scheduler for periodic jobs, authentication, storing and analyzing TradingView signals. If I couldn't ingest data, store it, and process it at set times, nothing else could get started.

I was greedy back then. I collected not only stock data but FX and crypto data as well, and went as far as letting AI make trading decisions and place orders. Looking back, I wasn't really building a product during that period. I was trying everything I could get my hands on to see what could be automated.

Rebuilding it as a tool where a person decides#

The big turning point was moving the center of gravity away from a system that tried to do everything, toward a stock screener, a curated stock pool, and a cockpit that supports manual decisions. I split data collection, analysis, the API, and the UI by role, and served everything from a single build and gateway, switching criteria and display per market: CN, JP, and US.

At the same time, I decided what not to hand over to the system. The cockpit doesn't place orders. I kept LLM trading decisions out of core, put deterministic GMMA and CCI signals and rules at the center of the analysis, and limited AI to a supporting role in a few places. It felt less like cutting features and more like getting clear, in my own head, about what any judgment on screen should be based on.

Designing to keep it running#

Picking stocks and showing them once was never going to be the end of it. They have to be evaluated on daily, hourly, and weekly bars, their prices and signals have to be monitored, and the results have to show up on stock detail pages and charts. Every time market data updates, evaluation has to move forward, workers have to process it, and the UI has to read the new state. Making each candidate's processing independent, so that a failure on one stock doesn't stop evaluation for the whole market, was an essential change in turning this into a product I could actually operate.

Making "when was this data finalized" part of the product#

The recent history is full of work on the JP and US stock universes, handling of newly listed and delisted stocks, rebuilding weekly data, and a fix so that daily and weekly bars are ingested only after the exchange's closing price is final. I learned the hard way that being able to fetch data isn't enough: which stock a value belongs to, and when it was finalized, is a premise of the analysis. If the weekly bars are inconsistent, any judgment that spans timeframes gets shaky; if you treat an unconfirmed close as final, every evaluation built on it is skewed. So I worked not only on generating data but also on the procedures for re-collecting it, detecting anomalies, and repairing it.

Showing how performance was measured#

At the end of October 2026, what I spend the most time on is rethinking how signal performance gets counted. I introduced a new aggregation based on whichever price condition is hit first, and invalidated signals are now treated as a separate state instead of being mixed in with wins and losses. I also excluded exit_warning from win/loss tallies and strategy tracking, because mixing warnings with trade outcomes changes how you read the performance in the first place.

To account for observations that cluster around the same stocks or time periods, I added confidence intervals using a cluster bootstrap. The admin screen shows the old and new methods side by side, so I can compare them without rushing to swap one metric for the other. I wrote "don't jump to conclusions about the cause" in my investigation notes because I didn't want to pin down why the numbers differed without evidence.

Honesty in the UI comes from the same place. It distinguishes AI confidence that hasn't been calibrated, risk that hasn't been assessed, and weekly verdicts that don't exist yet, and when a price is stale, it doesn't count the stock as "currently in the price range." States like no data, still fetching, and failed are shown as they are.

What I'm building now#

Looking back, what I've been building was never a machine that predicts stock prices. It's a platform that collects data from multiple markets, evaluates stocks with deterministic rules, delivers the results to screens where a person makes the decision, and keeps verifying the track record of the material those decisions rest on.

> First, I collected data and tried automation broadly. > Next, I shifted toward the stock pool and supporting manual decisions. > And now, I'm trying to show not just the signals, but how they're measured and how uncertain they are.

There are still uncommitted changes in the working tree, so what I've written here covers only the committed milestones. Where the next milestone will be, I haven't decided yet myself.

Contents