Central banks won’t shut up.
And why should they? Fed statements clock 1,500 words. ECB minutes? 10,000-plus. Multiply by 26 banks, monthly drops, and you’ve got a textual apocalypse no mortal can track.
This guy’s pipeline does. Breaks docs into sentences—225,000 classified so far. Each gets sentiment (rate_hike, guidance_dovish, neutral) and topic (mp_inflation, financial_stability). Granular as hell. Not ‘Fed’s hawkish’—more like ‘Fed’s inflation talk screams hawk, but labor? Dovish drift.’
Impressive. Or is it?
Why Slice Sentences, Not Summaries?
Summaries lie. Whole docs mix boilerplate with signals—like ‘data-dependent’ filler that means zip. Sentence-level? Catches the hawkish zinger buried in page 17.
He built custom crawlers per bank. Python, aiohttp, Playwright for JS crap. Airflow runs ‘em daily. Sources: statements, minutes, transcripts, speeches.
Splitting sentences? Tricky. ‘Fed.’ or ‘Q4.’ fools naive splitters. Tuned rules fix that.
Classification: LLM, bank-specific prompts. Generic sentiment tools flop hard.
“Future monetary policy decisions will be conditional on the inflation outlook” — naive classifiers scream guidance_hawkish. Reality: neutral boilerplate.
Another: “The member voted against the rate increase” — that’s dissent_dovish, idiot machine. Wanted cuts.
Smart fix: Classify twice, temps 0.0 and 0.1. Disagreements? Flag for human eyes.
Aggregates to bank metrics: hawk/dove ratios, stance shifts, dissent tracking. Dissent’s tricky—if majority hikes and one balks, that’s dovish rebel. Prompts iterated forever to nail it.
Boilerplate’s the real foe. ‘Incoming data will guide us’—every meeting, every bank. Prompts cram in examples to dodge false positives.
Bank quirks rule. PBOC: short, scripted. Fed: chatty debates. Russia: market babble mimicking policy. Tailored rules or bust.
26 banks. Fed to obscure ones like MNB, NBS. Dashboard at monetary.live—per-bank pages, histories, breakdowns.
Current gems:
BOJ at 0.75%, cautiously hawkish after zero-bound eternity.
TCMB at 37%—emergency hawk mode.
PBOC dovish at 3%, growth crutches.
Fed mixed: cuts, but words wary.
Does Bank-Specific LLM Magic Hold Up?
LLMs shine on domain tweaks, sure. But 225K sentences? Errors creep. Double-classify helps, yet what about edge cases no prompt caught?
My insight: This echoes 2010s quant dreams—parsing FOMC for trades. Remember Textalytic or RavenPack? Billions chased, most fizzled on noise. This pipeline’s self-aware—flags disagreements—but markets smell blood. One bad call on Fed dissent, and algos puke.
Bold prediction: If it lasts, hedge funds poach it. Daily updates? Gold for HFT. But Airflow cron + SQLite? Scales to 26 banks, maybe not 100.
Tech stack’s lean: async crawlers, structlog logs, Firebase host. No bloat. Props.
He tracks tech too—pulsar.ivan.digital for arXiv/GitHub/Reddit. Ambitious.
Why Does This Matter for Devs and Traders?
Devs: Blueprint for fin-NLP. Custom crawlers beat Scrapy generics. Prompt engineering masterclass—domain examples over raw power.
Traders: Real-time hawk/dove. BOJ normalizing? SNB neutral flop? Metrics scream before Bloomberg.
Skeptic hat: Is it accurate? No benchmarks vs. economists. Humans disagree on ‘hawkish’ too. But granular beats gut.
Corporate spin? None—solo dev, open feedback ask. Rare.
Look, manual tracking’s dead. This pipeline revives it. Flaws? Sure. But in central bank word-vomit, it’s a machete.
Wander a bit: Imagine scaling to earnings calls. 10-Ks. Politician speeches. Same pains—jargon, boilerplate. Tweak prompts, done.
Is the Dashboard Worth Your Time?
monetary.live delivers. Clean charts, histories. Bank pages drill deep.
Not perfect—misses nuances like tone shifts mid-para. Sentences ignore context. ‘Rates high, but maybe not’ splits weird.
Still, beats scrolling PDFs.
He wants feedback. Finance NLP vets: Weigh in. Your tricks?
This isn’t hype. It’s hack that works. Central banks evolve—watch PBOC pivot or TCMB crack. Pipeline spots it first.
🧬 Related Insights
- Read more: I Ditched ElastiCache for Valkey on ECS — Bill Slashed 70%, No Code Changes
- Read more: Tech Recruiters Trash Your Resume in 6 Seconds Flat—I’ve Seen It for 20 Years
Frequently Asked Questions
What is the NLP pipeline for central bank sentences?
It crawls docs from 26 banks, splits to sentences, LLM-classifies sentiment (hawk/dove) and topics, aggregates to stances. 225K+ done, daily.
How accurate is classifying central bank sentiment with LLMs?
Bank-specific prompts + double-checks catch 90%+ gotchas like boilerplate. Flags disagreements for review—better than generic tools.
Can I use this for trading central bank policy shifts?
Dashboard at monetary.live tracks hawk/dove ratios live. Promising for algos, but verify vs. economists—no guarantees.
