Research Note 06

What Is Left for a Retail Trader?

Unlike every other note in the series, this one runs the method backward: rather than starting from a published hypothesis, it starts from Klines, a detector already built from months of manual replay and case-by-case review of Binance volume spikes, and checks what it turned out to converge on only after it had a stable shape.

The literature check

Eleven papers are named directly, split across three questions. On why an anomalous volume spike predicts anything at all: Gervais, Kaniel & Mingelgrin (2001, Journal of Finance) on the volume-return premium in a different asset class and horizon; Kamps & Kleinberg (2018) and Xu & Livshits (2019) on defining and detecting crypto pump-and-dump events and the small, illiquid names they concentrate in; and — the load-bearing paper for timing — Dhawan & Putniņš (2023, Review of Finance), which finds the average pump peak arrives around eight minutes after the signal, with a late entry carrying negative expected value both theoretically and empirically. On exit management: López de Prado's triple-barrier method and meta-labeling, Sweeney's maximum adverse excursion framework, and Han, Zhou & Zhu's finding that a stop-loss overlay on momentum more than doubles the Sharpe ratio on a monthly-stock strategy, none of which was consulted before Klines' own exit structure — a monitoring window, then trailing exit, then hard timeout — had already converged on the same three-way race independently.

The statistics

On the larger, pre-patch walk-forward sample (93 tracked trades over roughly two months), the core "ALIVE" tier shows an 85% win rate with a mean return of +5.58% per trade, positive in every one of eleven consecutive weekly cohorts. A narrower "ROCKET" tier within it, defined by a stronger continuation move in the first five minutes, shows 100% win rate across 46 trades. A newer, more honestly timed log — using real detection timestamps instead of a hardcoded offset — covers only 67 signals over roughly two weeks, too short to serve as its own walk-forward verdict, and shows a materially weaker 63.6% ALIVE win rate; the rocket tier's 100% figure holds up on this smaller sample.

Two qualifications are stated directly rather than smoothed over. The rocket tier's definition partly overlaps with the outcome it predicts, so part of its 100% win rate is mechanical rather than purely predictive. And every number in the statistics is priced at a reference price recorded after a signal is confirmed — a price already reached after part of the move that makes the signal look good has happened. An earlier, more pessimistic backtest assumption flipped the same trade population from a large paper gain to a large paper loss purely by changing the fill assumption, and the newer, honestly timed log's weaker win rate points in the same direction, though on too small a sample to be conclusive.

Verdict

Klines is described plainly as semi-automatic: detection, tiering, and a short monitoring window are automated, while the decision to size into a specific trade remains a manual check layered on top. It currently runs in paper execution. The verdict states directly that nothing in the note should be read as a claim that Klines is fully automated or that its profitability is proven — its replay and paper-trade findings are walk-forward-stable on paper, not a guaranteed live edge, with execution validation still ongoing. A separate, structurally blind gap is also named: Klines is tuned to detect sudden volume spikes and cannot see a slow, low-drama continuation move in an already-liquid name, a different shape of event reserved for a separate, independently validated system.

Full note: GitHub ↗