Crypto Research Labs
Can a simple, published trading rule beat just buying a coin and holding it?
This project checks honestly. Every idea is written down and locked before it is tested,
and automated agents re-test old ideas as new market data arrives.
English ·
Русский ·
Español ·
Português ·
Türkçe ·
Tiếng Việt ·
中文
Not financial advice. This is a research project. It does not trade, manage money or sell signals.
A "pass" means "worth a closer look", not "this makes money".
New here? Start with this
| You want to… | Go to |
|---|
| Learn from zero, step by step | the course for beginners: 10 short lessons in 7 languages, numbers updated daily (на русском) |
| See today's market and every result, with plain-language explanations | the live dashboard (7 languages) |
| Understand the words (Sharpe, drawdown, p-value, funding…) | the glossary on the beginners' page, or docs/GETTING_STARTED.md |
| Know what was found | docs/FINDINGS.md |
| Which coins | BTC, ETH, SOL; for independent confirmation BNB, XRP, ADA, LTC, LINK, DOGE, XLM, TRX; Everscale (EVER) and Gram (GRAM, prev. Toncoin) from Gate.io. Any Binance spot pair can be added |
| Run it yourself or add your own idea | Quickstart, CONTRIBUTING.md or just post it in Discussions → Ideas |
| Ask a question | Discussions → Q&A |
The short answer so far: two ideas passed, and neither is a money machine. H-020 (stay out of Bitcoin after heavy exchange deposits) is a trading rule that is still being re-checked on live data. H-028 shows that a model can forecast how big tomorrow's move will be (risk), while the same model cannot forecast its direction (H-027 failed). Daily trend rules and
volatility targeting lower drawdowns, mostly because they spend less time in the market, but they
don't win by a margin that luck can't explain. Fast minute-level signals lose to fees, and the
famous calendar effects (Monday, turn of the month) did not survive after their papers were published.
What the dashboard shows
The site has five pages, each answering one question. The home page stays short: what has been found and what
the market looks like today; everything for beginners is one click away.
| Page | Question it answers |
|---|
| Home | What has been found, and what does the market look like today? |
| Market | Today's diagnosis, indicators grouped by trend, nervousness and money flows, the BTC chart, every coin |
| Research | Every hypothesis and its verdict (filter by verdict), which rule works on which coin, how we test |
| Today's model forecast | For each coin, a model's chance of a rise above fees and of a bigger move than usual, published every morning before 12:00 UTC and scored in public (H-027…H-030) |
| Digest | What happened this week, in plain words |
| New here? | How to read the site, a 10-lesson course and the glossary |
| About | How it runs, methodology, data sources and freshness |
| Market: what the market looks like today | Research: every idea and its verdict |
|---|
| |
| One hypothesis: numbers and equity curve | One coin: chart, numbers and what each tested rule says today |
|---|
| |
Every coin has its own page (BTC, ETH, SOL, BNB, XRP, ADA, LTC, LINK, DOGE, XLM, TRX, EVER, GRAM) with trend, volatility,
distance from its peak, what $100 invested a year ago is worth today and its correlation with Bitcoin, for example
Solana or
Everscale.
The home and Market pages open with a market diagnosis in plain words: one line (uptrend / mixed / downtrend)
and the reasons behind it (Bitcoin's trend, how many coins are rising, volatility, mood, stablecoin money,
the one rule that passed). Every coin page has the same kind of diagnosis for that coin.
Which rule works on which coin. A rule that works on Bitcoin may fail on Solana. Every rule tested in any
hypothesis is applied to every coin since 2022 (fees included), with a false-discovery check across the whole
table so lucky combinations are not mistaken for findings. The matrix is on the Research page and in
reports/COVERAGE.md; a ✅ on a new coin is only a candidate for a new pre-registered test.
Every Monday a language model retells the week in plain words, in all 7 languages: weekly digest (needs a free Groq or an Anthropic key, see Setup).
Tables can be sorted by clicking a column header. Hypotheses waiting for new data (H-022) show how many days
of it have been collected.
The design is high-contrast with depth (raised cards, a 3D verdict field, a 3D gauge) that never touches the text or the charts; motion turns off with the system "reduce motion" setting.
The site is rebuilt every morning by GitHub Actions, as soon as every source has published yesterday's data, from real Binance, OKX, DefiLlama,
alternative.me and BGeometrics data. Every page is fully rendered HTML, with hreflang language
versions, a sitemap and structured data, so it is readable without JavaScript and by search engines.
What we know so far
| ID | Hypothesis | Verdict |
|---|
| H-001 | Minute-level indicator ensemble predicts 1–4 h direction | ❌ Fail |
| H-002 | Inverted minute signals are profitable after fees | ❌ Fail |
| H-003 | Best daily trend rule (chosen 2018–2021) beats buy-and-hold out of sample | ◐ Weak |
| H-004 | Same, long/short futures | ◐ Weak on Sharpe, with drawdowns down to −87 % |
| H-005 | SMA50 long/flat beats buy-and-hold on 8 independent coins | ❌ Fail |
| H-006 | BTC long when stablecoin growth is above its 1-year median | ◐ Weak |
| H-007 | BTC out of the market when funding is in its top 20 % | ❌ Fail |
| H-008 | BTC out of the market when Fear & Greed ≥ 75 | ❌ Fail |
| H-009 | BTC long only after positive 5-day spot-ETF flow | ◐ Weak |
| H-010 | Stablecoin timing beats buy-and-hold on 8 independent coins | ❌ Fail |
| H-011 | BTC funding carry (spot long + perp short) | ❌ Fail (~1.5 %/yr recently) |
| H-012 | Volatility targeting (30 %/yr) on BTC, ETH, SOL | ◐ Weak: better Sharpe on 3/3, not significant |
| H-013 | Volatility targeting on 8 independent coins | ◐ Weak: 8/8 since 2022, 5/8 since 2019 (6 needed) |
| H-014 | 1-week momentum (Liu & Tsyvinski) on BTC, ETH, SOL | ◐ Weak: works on SOL, fails on BTC |
| H-015 | Monday effect in BTC | ❌ Fail |
| H-016 | Turn-of-the-month effect in BTC | ❌ Fail |
| H-017 | Altcoins only while BTC is above its 200-day average | ❌ Fail: 7/8 since 2022, 1/8 since 2019 |
| H-018 | Everscale (EVER): SMA50 trend rule | ◐ Weak: −10 %/yr vs −70 %/yr holding, not significant |
| H-019 | Everscale (EVER): only while BTC is above its 200-day average | ◐ Weak |
| H-023 | Gram (GRAM, prev. Toncoin): SMA50 trend rule | ◐ Weak: +17 %/yr vs −9 %/yr holding since mid-2023, drawdown −56 % vs −85 %, p = 0.11 |
The live table is in reports/RESULTS.md and on the dashboard.
How it works
flowchart LR
A[Data steward<br/>daily] -->|shared cache| B[Regime watch<br/>daily]
A --> C[Hypothesis runner<br/>weekly]
C --> D[results/ + reports/]
B --> E[Telegram alert<br/>only on change]
C --> E
D --> F[Dashboard<br/>GitHub Pages]
B --> F
G[Proposer<br/>monthly, optional LLM] -->|draft PR| H[You review<br/>and lock]
H --> C
| Agent | When | What it does |
|---|
| Data steward | daily | Updates one shared data cache and checks it for gaps, stale series, duplicates and outliers |
| Regime watch | daily | Tracks funding carry, stablecoin growth, sentiment, trend and volatility; alerts only when a reading changes band |
| Hypothesis runner | weekly | Re-tests every due hypothesis the same way, stores JSON results, alerts when a verdict changes |
| Digest | weekly | A short summary of data, regime and results (optionally rewritten by Claude) |
| Dashboard | daily | Builds the multilingual site and publishes it to GitHub Pages |
| Proposer (optional, LLM) | monthly | Claude drafts new hypotheses; they are validated, de-duplicated and opened as a PR for a human to review |
The LLM can't run tests or change verdicts. A human locks every draft (SHA-256) before it is
ever tested. Why it's built this way: docs/INEFFICIENCIES.md.
Quickstart
You need Python 3.10+ and git.
git clone https://github.com/f1devbin/Crypto-Research-Labs
cd Crypto-Research-Labs
python -m venv .venv && . .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest -q # self-tests: finds planted effects, rejects noise, no look-ahead
crl registry check # every hypothesis valid, locked and unique
crl data refresh # download market data into ./data (first run: ~15 minutes)
crl run --force # re-test all hypotheses
crl watch # today's market regime -> reports/regime.md
crl scan # every tested rule on every coin -> reports/COVERAGE.md
crl site # build the dashboard into ./_site (open _site/index.html)
Run your own copy on GitHub
- Fork the repository.
- Settings → Pages → Source: GitHub Actions.
- Actions → research-agents → Run workflow. The first run takes about 20 minutes.
After that it runs on its own every day.
Optional settings (Settings → Secrets and variables → Actions). A step-by-step guide in 7 languages, with how to get the bot token, your chat ID and an AI model key, is on the dashboard: Setup · Настройка.
| Name | Kind | What it does |
|---|
TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID | secret | sends alerts to Telegram |
TELEGRAM_LANG | variable | language of Telegram alerts: en (default), ru, es, pt, tr, vi, zh |
ANTHROPIC_API_KEY | secret | turns on the proposer and the LLM digest |
LLM_API_KEY | secret | free alternative: Groq (default) or any OpenAI-compatible key; wins over ANTHROPIC_API_KEY |
CRL_LLM_BASE_URL | variable | OpenAI-compatible endpoint (default https://api.groq.com/openai/v1); the model is picked from its /models list unless CRL_LLM_MODEL is set |
CRL_LLM_MODEL | variable | overrides the model (default claude-sonnet-5-5) |
GOOGLE_SITE_VERIFICATION | variable | Search Console HTML-tag code, added to every page |
CRL_SITE_URL | variable | dashboard URL if you use a custom domain |
For the proposer to open PRs, turn on Settings → Actions → General → Allow GitHub Actions to
create and approve pull requests.
GitHub's runners are in the US, where fapi.binance.com answers HTTP 451. Recent funding
therefore comes from OKX. Binance's historical archives aren't blocked.
Project layout
src/crl/data/ downloading, caching and checking market data
src/crl/engine/ trading rules, backtest, statistics, protocol (spec -> verdict)
src/crl/agents/ data steward, regime watch, hypothesis runner
src/crl/llm/ optional Claude layer: proposer, digest
src/crl/site*.py dashboard builder; src/crl/i18n.py holds the translations
hypotheses/ one YAML file per idea, SHA-256-locked
results/ reports/ written by the agents
config/watch.yaml what regime watch monitors
docs/ getting started, methodology, findings, workflows
Lessons kept from the past
- One clock: UTC. Mixing UTC and UTC+3 once created a 30-minute blind window.
- Stops live on the exchange. A stop held only in a bot's memory disappears when the bot goes down.
- Prove the edge before building infrastructure. A year of engineering went into a bot whose signals turned out worse than random.
Contributing
New hypotheses, translations and fixes are welcome. See CONTRIBUTING.md.
License
MIT. See LICENSE.
