FinTechFeed

A finance research agent that reads the market's chatter for you.

FinTechFeed pulls headlines and posts about your watchlist from free, no-key public sources, scores each with a finance-tuned sentiment model, and rolls it all up into a ranked, cited daily sentiment digest — in your terminal, as Markdown, or as JSON for a downstream pipeline.

No API keys. No paid data. pip install and run.

fintechfeed digest --tickers NVDA,TSLA,BTC

Give an analyst's morning read — scan the tape, gauge the mood, cite the sources — to a 30-second command.

---

Why I built it

I trade a live multi-asset book and keep a structured research journal, and the first 30 minutes of every session is the same ritual: scan the tape, read the headlines, gauge whether the mood on a name is turning. FinTechFeed automates the gathering and scoring of that ritual so the human part — the thesis — starts from an evidence base instead of a blank page.

It's deliberately built around three ideas that matter in real research work:

1. Signal over vibes. Sentiment is weighted by source trust (curated financial press counts more than a retail thread) and every score is backed by clickable evidence, so you can audit any number.
2. Resilience over completeness. Each data source is an independent, swappable channel. If one rate-limits or breaks, the digest still runs on whatever else responded — it never returns partial garbage.
3. Zero friction. It runs on a fresh clone with no keys, so the output is reproducible by anyone.

Demo

Real output from fintechfeed digest (Markdown format, abridged):

Ticker	Signal	Score	Mentions	Sources
AAPL	🟢 Bullish	+0.29	44	hackernews: 23, reddit: 2, yahoo_rss: 19
NVDA	🟢 Bullish	+0.19	48	edgar: 3, hackernews: 20, reddit: 3, yahoo_rss: 22
TSLA	🟢 Bullish	+0.18	44	edgar: 1, hackernews: 21, reddit: 1, yahoo_rss: 21
BTC	🟡 Neutral	+0.03	51	edgar: 1, hackernews: 21, reddit: 3, yahoo_rss: 26

Evidence — TSLA (Bullish, +0.18)

• +0.97 Tesla's new SUV could be its smartest profit play yet — yahoo_rss
• +0.91 Tesla (TSLA) Lands $9 Billion In Megapack Orders And Expands Home Energy Push — yahoo_rss

(SEC 8-K filings score neutral by default, so edgar shows up in the source counts above but rarely in the top-sentiment evidence — it moves the needle only when a filing signals something material like an impairment or delisting.)

A full sample run is committed at docs/sample_digest.md (and docs/sample_digest.json).

Quickstart

git clone https://github.com/anishkhetani/fintechfeed.git
cd fintechfeed
pip install -e .

# Check what works in your environment (probes every source live)
fintechfeed doctor

# Run a digest on the default watchlist
fintechfeed digest

# Focus on a few names, write Markdown to a file
fintechfeed digest --tickers NVDA,BTC,ETH --format markdown --out today.md

# Machine-readable output for a pipeline
fintechfeed digest --format json --out today.json

# Every run is saved locally; see how the mood has moved over time
fintechfeed history                 # matrix of recent runs, all tickers
fintechfeed history --ticker NVDA   # one name's day-by-day trend

Requires Python 3.10+.

How it works

                 ┌── yahoo_rss ──┐
                 ├── edgar ──────┤
   watchlist  →  ├── reddit ─────┤ →  ticker resolver  →  sentiment  →  aggregate  →  digest
                 └── hackernews ─┘     ($TAG + aliases)    (finance      (weighted     (terminal /
                                                            VADER)        mean+label)   md / json)

1. Sources each fetch market chatter and normalise it to a common Item shape. They're registered in a table, so config alone turns them on/off. A shared HTTP layer adds timeouts and polite retry/backoff (honouring Retry-After).
2. Ticker resolution maps each item to your watchlist via $CASHTAG detection, whole-word alias matching ("Apple" → AAPL, but not "applesauce"), and per-feed hints, with a stop-word guard against common words like AI/CEO.
3. Sentiment uses VADER as a baseline, extended with a finance lexicon so market language is scored correctly — beat, upgrade, and guidance raised read bullish; miss, downgrade, and guidance cut read bearish.
4. Aggregation computes a source-weighted mean compound score per ticker, buckets it into Bullish / Neutral / Bearish, and attaches the most opinionated evidence.
5. History appends each run to a local file, so the next digest shows a day-over-day delta and flags a ticker as turning when its label flips or its score moves past a threshold — the "mood is turning" signal.

Mood is turning: day-over-day deltas

Because every run is recorded, the digest tells you not just where sentiment is but which way it's moving — the part an analyst actually reacts to. From the second day onward a Δ 1d column appears and turning names are called out:

| Ticker | Signal | Score | Δ 1d | Mentions | Sources |
| ------ | ------ | ----: | ---: | -------: | ------- |
| **NVDA** | 🟢 Bullish | +0.18 | ▲ +0.53 🔄 | 42 | edgar: 5, hackernews: 15, yahoo_rss: 22 |
| **TSLA** | 🟢 Bullish | +0.19 | ▲ +0.14 🔄 | 38 | edgar: 2, hackernews: 15, yahoo_rss: 21 |
| **BTC**  | 🟡 Neutral | +0.01 | → +0.00    | 43 | edgar: 1, hackernews: 16, yahoo_rss: 26 |

> 🔄 Mood turning: NVDA (Bearish→Bullish), TSLA (Neutral→Bullish)

This is also where the edgar channel earns its weight: a fresh material 8-K (an impairment, a delisting, a bankruptcy) can be the thing that tips a name's day-over-day score and trips the turning flag.

Sources (all free, no API key)

Source	What it gives you	Notes
yahoo_rss	Per-ticker financial-press headlines	Highest signal; items arrive pre-tagged
edgar	SEC 8-K material-event filings, per ticker	Primary source (highest trust); resolves ticker→CIK, reads the 8-K item codes. Neutral by default, decisive on impairment/delisting/bankruptcy
reddit	Retail sentiment from configurable subreddits	Public RSS feeds; Reddit rate-limits aggressively by IP
hackernews	Tech/crypto stories via the free Algolia API	Complements the finance press

SEC etiquette: EDGAR asks API clients to send a contact in the User-Agent. FinTechFeed ships a working default so it runs on a fresh clone; set your own via edgar.user_agent in config or the SEC_EDGAR_USER_AGENT env var.

Adding a source is ~30 lines: subclass Source, implement fetch(), and add one line to the registry.

Configuration

Copy config.example.yaml to config.yaml (git-ignored) and edit your watchlist, sources, source-trust weights, and sentiment thresholds. Everything has sensible built-in defaults, so config is optional.

watchlist:
  NVDA: ["Nvidia"]
  BTC: ["Bitcoin", "BTC-USD"]
sentiment:
  source_weights: { yahoo_rss: 1.0, edgar: 1.2, reddit: 0.6, hackernews: 0.8 }
  min_mentions: 2
  turning_delta: 0.1        # day-over-day move that trips the "turning" flag
history:
  enabled: true             # append each run; disable, or use `digest --no-save`
  path: ".fintechfeed/history.jsonl"

Optional: LLM desk brief

If you set llm.enabled: true, install the extra (pip install fintechfeed[llm]), and export ANTHROPIC_API_KEY, FinTechFeed will ask Claude to write a short analyst-style narrative grounded strictly in the evidence it already gathered (it is prompted not to invent prices or events). It's fully opt-in and degrades to no-narrative if anything is missing.

Design notes & honest limitations

• Headline sentiment is a proxy, not a price target. Lexicon models read the tone of a headline, which can diverge from a stock's actual outlook. FinTechFeed surfaces and cites evidence to reason from; it is a research aid, not a trading signal, and nothing here is financial advice.
• Reddit is best-effort. Reddit rate-limits public feeds aggressively; the resilient design treats a throttled source as a skipped channel rather than a failure.
• Per-ticker feeds include adjacent market news, so a ticker's bucket can contain sector-wide items — intentional, since sector tone moves names.
• Day-over-day deltas are only as regular as your runs. The baseline is the most recent run from a prior calendar day; if you skip days, the "1d" delta simply spans the gap. The first run on a fresh clone has nothing to compare to, so no deltas show until the second day.

Roadmap

• SEC EDGAR 8-K / filing sentiment channel
• Sentiment history + day-over-day deltas (a "mood is turning" signal)
• StockTwits and a FinTwit list channel
• Optional local FinBERT scorer as an alternative to the lexicon

Development

pip install -e ".[dev]"
pytest          # unit tests (offline; sources are faked)
ruff check .    # lint

Acknowledgements

The initial spark for this project came from studying Agent-Reach by @Panniantong — a CLI that gives AI agents multi-platform internet access. Its "treat each platform as an independent, swappable channel" framing shaped how FinTechFeed's source layer is organised.

To be clear about what is and isn't borrowed: FinTechFeed is an independent, from-scratch implementation in a different problem domain (finance sentiment research, not general agent tooling). No code is copied from Agent-Reach — only the high-level channel-architecture idea was an influence. Credit for that idea belongs to its author; any bugs here are my own.

Also built on the excellent open-source work of VADER (sentiment baseline), feedparser, and rich.

License

MIT © Anish Khetani — see LICENSE.