ShumiShumi

Studies

Proof of Work.

Benchmarks, backtests, and signal discoveries. Methodology included.

If a model underperformed, the autopsy is here. If a result held up, the notes, caveats, and method sit next to it.

5 published · 5 in review

Benchmark · Mar 11, 2026

Crypto AI Benchmark v2

6 tools · 9 prompts · 270 weighted scores

Research notes

Source material behind the episodes. Data, methodology, and limitations.

Study

May 13, 2026

3 min read · 4 min watch

How Shumi Tracked the Spring 2026 Rally

Update · September 28, 2026 The regime engine behind this study did not hold up in later forward testing, and its entries are no longer promoted. The numbers below are what it recorded in spring 2026, kept as that record.

The spring 2026 rally was the first full window Shumi tracked end-to-end. Twenty-three outcomes closed inside it, fourteen of them winners, and the winners summed +224%. None hit the 30% stop. (Numbers here reflect the current ledger. The video walkthrough used an earlier snapshot of the same window.) Here's what Shumi ran, what it skipped, and what's still unfinished.

Study

Mar 25, 2026

5 min watch

400 Experiments. 3 Survived.

Update · September 28, 2026 A stricter retest on September 7, 2026, with a random baseline that trades the same coins on the same days the way the engines do, found the surviving engines no better than chance, and they lost money on 2026 data. The method story below stands; the "3 survived" headline does not.

400+ strategy variants tested across three research tracks and five assets from 2021 to 2025. Three survived walk-forward validation. This study covers the mean reversion system: 77 out-of-sample outcomes with entry and exit prices sourced from Binance USDT perpetual data.

Source note

Mar 10, 2026

2 min read

Walk-Forward Performance, 2024-2025

Update · September 28, 2026 A stricter retest, which keeps the engine's trades across many coins together the way it actually trades, found these results no better than chance, and the engine lost money on 2026 data. The numbers below are the 2024-2025 test as first published, kept as that record.

Two years of out-of-sample outcomes with actual Binance funding costs applied per entry. 26 wins out of 29 outcomes. Weak spots left in.

In review

Drafts and recurring briefs moving through editorial.

Study

Draft

Forward Test Results: 3 Months Live

Follow-up to '400 Experiments. 3 Survived.' The surviving systems met 2026 data and a stricter retest, and did not hold up: no better than chance, and a loss on 2026 data. What went wrong, what the earlier tests missed, and how results are reported from here.

Research note

Draft

Sentiment Engine Drift Audit

Tracking how narrative momentum decays after the initial news event. Which categories (L1s, DeFi, memes) mean-revert fastest, and when sentiment flips from signal to noise.

Research note

Draft

Deep Research Source Quality

Source-by-source audit of the long-form research mode. Measuring freshness decay, contradiction rates between providers, and retrieval accuracy against known-answer queries.

Field brief

Series

Funding State Briefs

Short recurring posts on crowded funding, carry opportunities, and the setups that actually deserve attention that week.

Case study

Draft

Signal Receipt Case Studies

Post-mortem case studies on live signals: entry conditions, execution timing, and where risk allocation could improve with hindsight.