> PelvicOarfish: Using Data Compression Algorithms for Stock Market Prediction

Badros, phil kos, same difference.

This is the DCC '94 paper from the Duke group (Azhar, Badros, Glodjo, Kao, Reif). Short verdict: the theoretical framing is legitimate and was ahead of its time, but the empirical section doesn't support the claim, and the headline 55.06% is probably not evidence of prediction at all.

**The central tension.** Compression implies prediction only for stationary ergodic sources. Sections 1.3 and 4.1 argue at length that markets are non-stationary, multimodal, and fractal. Then the actual experiment is LZ on uniformly quantized log returns, which is the stationary-ergodic assumption they just spent three pages rejecting. Everything built to handle non-stationarity (dynamic memory windows, DCVQ, TSVQ) is promissory: "we are investigating," "shall report in near future."

**Why the 55% is suspect:**

- **No baseline.** Major indices drift up on roughly 52–54% of days. Their DowJones (51.10) and Snp500 (51.42) are *below* always-guess-up. For the two deepest markets in the sample, LZ loses to a constant.
- **The pattern in the table is a liquidity tell.** The top scores are Spain 62.63, Finland 61.85, Portugal 61.28, Indonesia 60.03. Those were the thinnest markets of that era. Stale prices and non-synchronous quotes produce serial correlation that a compressor eats happily and a trader can't touch.
- **Self-reported contamination.** They admit including flat periods with no price change. Zero-return days scored as correct guesses inflate the number directly, and they wave it off as "very few instances" without quantifying it.
- **Undisclosed abstention.** "If we are at a leaf, we do not guess" means coverage is unknown, so the denominator is unknown.
- **No significance test, no multiple-comparison correction across 22 markets, no transaction costs, no P&L.** Hit rate without conditional magnitude says nothing about whether the strategy makes money.

**Conceptual slip in §1.3.** They conflate stable Lévy distributions with fractional Brownian motion, saying non-integer α yields fBm. α is a tail index, H is a memory exponent. Those are different generalizations of Brownian motion. Related: the R/S-based memory window rests on Hurst estimation, and Lo (1991) showed R/S is severely biased by short-range dependence, wiping out most claimed long memory in returns. They cite Lo, just for a different point.

**What's still worth taking.** The compression-as-structure-detector idea survives, and they had the tool in hand without using it. The clean test isn't "predict direction and count wins," it's: symbolize the return series, measure LZ code length, compare against shuffled and phase-randomized surrogates of the same series. If the real series doesn't compress meaningfully better than its surrogates, there's no exploitable structure, and you get a significance test for free instead of arguing about 55 versus 52. Their §2.3 multispectral idea is also just cross-sectional factor structure in compression vocabulary, and the modern successors to §3 are variable-order Markov models and context-tree weighting, which handle the adaptive-order problem far better than hand-rolled tree splitting.