sigstar

Collect market, on-chain, news and macro data, and explain, forecast and alert on crypto price moves from every angle. In development. Python, Apache-2.0…

What it does

Prices alone don't say why a market moved. sigstar collects the data that does, organises it in one place, and analyses each move from every angle, on your own machine.

uv tool install git+https://github.com/Bit-Primitive/sigstar
sigstar init        # writes sigstar.yml: assets, sources, rules
sigstar run         # collect, build, alerts and the daily brief
sigstar explain ETH --top 3

• explain takes the biggest moves (or a date you give) and ranks what else was unusual at the time: intraday range, volume, funding and open interest, the price gap between exchanges, Fear & Greed, news volume and tone with the day's headlines, macro changes, and what the other assets did. Each driver comes with its size against the previous 30 days. Coinciding is not causing, and sigstar says so every time.
• forecast forecasts price, volatility and volume with ETS and ARIMA, backtested over rolling origins against a naive forecast, and shows the naive one when no model beats it by at least 2%.
• relationships reports correlations, betas between assets, lead and lag, Granger tests corrected for testing many at once, and volatility regimes.
• anomalies lists every feature far from its own recent history.
• alerts evaluates rules you write in YAML ("funding above its 95th percentile while news tone falls") and logs each alert with its evidence, optionally to a webhook.
• report writes a self-contained daily brief in Markdown and HTML, with charts and no external scripts.

Every angle, from public sources

Every source is used through its public API or feed, within its terms: sigstar identifies itself, keeps to each host's published rate, and obeys robots.txt. Sources whose terms forbid this kind of use are left out, and the project's SOURCES.md says which and why. News is kept as headline, link and a short snippet, never full articles.

Several exchanges license their public data for personal use only, so in its default personal mode keep what sigstar collects and the reports it writes to yourself.

Commercial-safe mode

Set sources.licence_mode: commercial and sigstar switches off every source whose terms restrict commercial use or don't clearly allow it, including the exchanges. It runs on sources that do allow it (the Fear & Greed Index with its required credit, GDELT news, public-domain FRED series) and on prices derived from on-chain DEX swaps by chainstar, which carry no third-party terms. Every command and report states the mode and the sources whose data was actually used.

Honest about thin data

On-chain figures are compared only with days on which the same chains were loaded to a similar extent. When that isn't possible, sigstar leaves the figure out of explanations, anomalies and alerts, and says why, rather than reporting a jump that only reflects more data being loaded.

How it is built

dlt collects, incrementally; dbt organises it into tested tables, ending in one feature table per asset and day; DuckDB stores it; statsmodels does the statistics.

Ground rules

sigstar produces analytics and signals you define. It is not financial advice, and every report and alert says so.

Status

In development: Bitcoin, Ethereum and Solana, on Linux. Macro needs a FRED key; GDELT limits how often it can be asked, so news history fills in over several runs.