“The strategy does not work and cannot be made to work by tuning parameters.”

apes-signal, final-report.md

What this looks like for me

Most of my quant work has been research platforms rather than live trading. Screening pipelines, backtesting frameworks, signal studies. The parts I spend the most time on are not the strategies themselves but the scaffolding around them: how data snapshots are hashed, how in-sample and out-of-sample splits are enforced, how a run is traceable back to the commit that produced it, how the pipeline reports confidence alongside the headline number.

A fair share of these projects end with a negative result. InsiderSignalResearch was shelved after a full sprint over 4,143 SEC insider filings. apes-signal ran a ratchet optimiser to a documented ceiling. In both cases the reason I kept the work around is the pipeline — the same instrumentation survives even when a specific hypothesis doesn’t.

I’m more comfortable building on fundamentals and structured data than on pure price-action edge, and more comfortable with transparent scoring than with black-box models. That’s a preference rather than a hard rule.

Projects that back this

  • edenfintech-scanner-python — fundamental screener and scoring pipeline with an agent-level information barrier for second-opinion review. Stdlib-only core.
  • InsiderSignalResearch — IS/OOS research programme over SEC Form 3/4/5 insider-buy signals. 4,143 signals, 27 logged experiments, shelved with a written postmortem.
  • ftr_strategy_backtesting — three systematic strategy families on a shared metrics spine. Confidence-adjusted expectancy, no-lookahead HTF filtering, walk-forward with overfit flags.
  • apes-signal — bar-by-bar Pine-to-Python port of a momentum indicator, wrapped in an iterative parameter-search loop. Publicly archived negative result.

Decisions that shaped how I do it

What I’m usually asked to do

  • Set up a research pipeline from scratch: ingestion, snapshot hashing, config files, CLI
  • Review an existing backtest for look-ahead, overfitting, or unrealistic assumptions
  • Port a TradingView / Pine indicator into Python with its original semantics preserved
  • Structure a research sprint so the stop criteria and postmortem are part of the plan