The RS Filter Problem Wasn't the Formula — It Was a Skewed Benchmark

May 22, 2026

auto-tradingRS-filterregimestock-auto-trade

Summary

I suspected that using the KOSPI200 ETF as the RS benchmark was unfairly filtering out stocks because of mega-cap concentration. Recalculating with a universe-weighted benchmark produced the opposite of what I expected: market-cap weighted universe RS came out higher than the ETF. The problem isn’t the ETF — it’s a skew that exists in every commonly used benchmark.

I Thought the ETF Was the Problem

The auto-trading system measures a stock’s relative strength by comparing its RS against the KOSPI200 ETF. The suspicion: the ETF is heavily weighted toward Samsung Electronics, SK Hynix, and similar mega-caps, so comparing individual stocks against it might knock out perfectly good candidates unfairly.

The natural fix seemed to be switching to the universe of 40 stocks instead. Intuitively, that should give a fairer comparison.

The Data Reversed the Intuition

Market-cap weighted universe RS came out at +36.57% — higher than the ETF at +27.34%. The universe itself is concentrated in a few mega-caps just as much as the ETF.

Switching the Benchmark Doesn’t Fix the Skew

This is the key insight. The problem isn’t “we used the ETF.” If a benchmark is skewed toward large caps, the RS filter will be distorted regardless of whether you use the ETF or the universe. A biased benchmark produces a biased filter.

That’s why changing the benchmark right now is risky. Simple average lowers the bar and floods the filter with candidates. Market-cap weighted makes the filter even stricter. Candidate pool weighted hollows out the filter’s purpose. Every alternative trades one distortion for another.

Build the Verification Tool First

Rather than fixing the benchmark, I built a tool to test whether the current filter is making good decisions. The plan: track every stock rejected by the RS filter and record its D+1, D+3, and D+5 performance.

Was the filter blocking missed gains? Or was it preventing chases into overextended stocks? That question needs data, not intuition.

Next observation: once data accumulates, the first validation metric will be whether D+5 average returns for RS-rejected stocks run higher or lower than held positions.

What Matters in Auto-Trading

The goal isn’t to buy more. It’s to have a structure where every buy decision — and every non-buy decision — can be verified with data. Today added one more layer to that structure.