Research Note 00

Can You Trade Crypto Profitably by Formula?

The series opens with a starting question rather than a formal hypothesis: if hundreds of ready-made trading tools already exist on public marketplaces, why would building or copying one produce any advantage. Oleg Arefev audited roughly 200 pre-built strategies from a marketplace referred to in the note as a "CMC Skill Hub," shortlisted around 15, and reverse-engineered their underlying mechanics — open interest, funding, flow imbalance, liquidations, ATR, holder concentration. The audit's conclusion was that almost none relied on genuinely proprietary data; most recombined the same public metrics anyone can compute for free on Binance data. One specific "copy-paste bot" was reverse-engineered in detail and turned out to be exactly that: standard free metrics wrapped in a polished paid interface.

Before this question was even posed, Klines already existed as an independent, real-time event detector, built from months of manual replay and case-by-case review of Binance volume spikes. Its own finding shaped everything that followed: detecting an abnormal volume event turned out to be considerably easier than judging, at the moment it happens, whether the move will continue, exhaust, or reverse. That gap is what sent the project into the academic literature rather than toward adding more indicators to the existing detector.

Four independent hypothesis classes were selected for testing under one fixed protocol, applied without exception: the published claim is taken at face value first, its mechanism reconstructed, then replicated on real Binance data, priced gross, then net of realistic execution costs, then tested for statistical significance. The four candidates are momentum, trend-following, funding carry, and reversal. Grid trading was set apart from this protocol entirely, treated as a question of execution economics rather than a market pattern, because its apparent profitability was judged likely to depend on fee tier and account scale rather than on a predictive edge.

A rule was fixed in advance, before any of the four hypotheses were tested: at most two independently published implementations are tested per class. If both fail, the class is closed, and no third variant is tried to rescue it. Scalping and high-frequency trading were excluded from the series altogether on the grounds that retail infrastructure cannot compete with millisecond-latency counterparties — a structural disadvantage, not a skill gap. More than twenty to thirty academic papers were reviewed to build this protocol, spanning momentum, reversal, carry, market microstructure, pump detection, and meta-labeling, with the explicit caveat that a meaningful share of these studies end their samples before 2023 and therefore had to be re-checked against current Binance data rather than taken on trust.

This note carries no verdict of its own in the sense the later notes do. It sets the discipline every subsequent note in the series — grid, momentum, trend, carry, reversal, and Klines' own retrospective check — is held to, and states plainly, before a single result is in, that the series expects to close more hypotheses than it confirms.

Full note: GitHub ↗