Tools for my dad's research
My father has spent decades studying market time cycles. I build the tooling around his work: automation, checks and indicators he can verify on his own screen.
Plumbing for a craft
The method is his and it stays his. What I can do is remove the friction around it: repetitive screen work, hard-to-compare outputs and indicators he has to draw by hand.
Three kinds of help
- Automating a desktop analysis program. The software he trusts is a desktop application with no API. I built agents that drive its interface step by step to produce the monthly forecast chart, so the run happens on its own.
- Honest scoring. A forecasting harness that checks any prediction method out-of-sample, walk-forward, against a random-walk baseline. His own question was simple: does it get the direction between two timestamps right? That question became the metric.
- Indicators on his chart. A growing library of TradingView indicators written in Pine Script from his written specifications. He verifies each one on TradingView before it counts as done.
A good spec is the product
The work runs through documents: a method written up, turned into a spec, turned into code, checked by the person who knows it best. The lesson is that building for an expert is mostly about making their judgement the final test, and never replacing it.
What it uses
Python, k-nearest-neighbour and ridge models for experiments, Pine Script for charts, a small MCP server so agents can read TradingView, and screen-driving automation for the desktop program.