Teaching a local neural net to read political tape: vectors, trades, and how you can help

Political markets move on stories. Machines only “read” stories if we turn them into numbers. This week 3DN stood up a small, local machine-learning spine for PolitiCap: every English desk article is embedded into a Postgres vector database, then joined to real executed trades on the civic tape. The goal is not a casino crystal ball. It is a local, inspectable model that may eventually help forecast short political swings on virtual tickers — if the community generates enough honest trading activity to learn from.

Sign up and trade: politicap.eu — free account, dibs (virtual DutchBud credits) on a closed-loop ledger. More trades after real news events = better labels for the net.

What we built

The pipeline is deliberately boring and fleet-local:

1. EN article on politicap.eu (pretty-linked tickers in the body).
2. Embedder sentence-transformers/all-MiniLM-L6-v2 → 384-dimensional vectors (same family as our fleet RAG).
3. Postgres database politicap_ml with pgvector (HNSW cosine indexes).
4. ETL builds article_symbol_windows: for each article×symbol×horizon, pull executed orders after publish time.

Vectors: what “similarity” means

Each English post becomes a unit vector e in ℝ³⁸⁴. Chunks of the body get their own vectors for finer retrieval. Nearby points in that space tend to share themes (AI race, TRUMPD geopolitics, Dutch cabinet wire).

Cosine similarity between two L2-normalised vectors u and v collapses to a plain dot product:

sim(u, v) = (u · v) / (‖u‖ ‖v‖) = uᵀ v

That is useful for “stories like this one.” It is not by itself a price forecast. Trades enter through a different door.

How trades attach to embeddings

We never stuff order IDs into the vector. The join is an event spine:

• article_tickers.symbol — who the story is about (from data-symbol pretty-links).
• articles.published_at = t₀ — when the news event starts (UTC).
• Horizons H ∈ {5 min, 30 min, 2 h, 1 day} — only activity after t₀ counts as reaction.

For last price P and executed volume V in the window:

• Simple return: r_H = P(t₀+H) / P(t₀) − 1
• Log return: ℓ_H = ln P(t₀+H) − ln P(t₀)
• Volume shock: z_H = (V_after − V_prev) / max(V_prev, 1) comparing the post-event window to the equal-length window just before t₀

Those numbers land in article_symbol_windows next to the embedding. That table is what a neural net (or a boring gradient booster) actually trains on.

A predictive model — without the hype

A training row looks like:

x = [ e_doc ‖ f_pre(s, t₀) ‖ 1_s ]
ŷ_H = g_θ(x) ≈ r_H or sign(r_H) or “volume spike”

Here e_doc is the frozen article embedding, f_pre are market features before the news (so we do not leak the answer), and 1_s tags the symbol. Start with ridge or gradient boosting on frozen MiniLM features. Only then a small neural net. Walk-forward splits by time. Beat a dumb baseline first.

Disclaimer: PolitiCap is a civic game with virtual DutchBud credits — not licensed securities advice, not a prediction market for fiat cash-out. A “successful” model means better description of virtual tape dynamics, not a claim you can bank on in the real world.

First map: top 10 symbols by news-linked 1-day volume

After embedding 203 English posts and joining ~247k executed fills, these symbols carried the most post-article volume in 1-day event windows (summed across tagged stories — overlapping AI posts can share the same day’s tape):

#	Symbol	1d volume (sum)	Top news drivers (by window volume)
1	AI	867 247	Let there be Light; China AI slowdown “Cold War playbook”
2	TRUMPD	757 480	GROKAI learns candles; TRUMPD Iran-plan tape; Taiwan survey
3	GROKAI	627 238	Let there be Light; China AI slowdown; AI Index free-float soak
4	MUSKE	606 844	FBI Grassley emails; MUSKE/MENGW IPO; candles battle note
5	CHATGPT	579 296	Let there be Light; China AI slowdown; free-float soak
6	METAAI	493 536	Same Battle / AI-index cluster
7	GEMINAI	470 226	Same Battle / AI-index cluster
8	DEEPSEK	410 550	Same Battle / AI-index cluster
9	LIANGW	236 176	GROKAI learns candles (primary)
10	ALTMS	204 734	GROKAI learns candles; Grok 4.7 desk note

Pattern: Battle of the AIs listings and TRUMPD/MUSKE geopolitics dominate labeled volume. That is exactly the kind of repeated event structure a local model can learn — if humans keep showing up on the book when the desk drops a wire.

How people make the model better

1. Trade after real news — your fills become (symbol, t₀, H) labels. Thin books teach noise.
2. Use the ticker links — pretty-links are the join keys into the tape.
3. Do not game empty symbols — wash volume without news confounds the windows.
4. Stay on the closed loop — dibs / DutchBud virtual bank language keeps the civic market honest: fintech spine, not casino cash-out.

Create a PolitiCap account → play the civic market, leave a trail of trades the lab can learn from.

Stack notes (for the curious)

• DB: politicap_ml on fleet Postgres with pgvector
• Embed on publish: desk skill runs embed_en.py --only-id for every EN post
• Windows ETL: executed orders (status=E) from the PolitiCap MySQL book
• Next: walk-forward baselines, then a small net — still local, still ours

3DN builds the rails: sovereign compute habits, open tooling, and products that stay under our fingers. A predictive civic tape is one more reason the stack — hosting, bank, marketplace, PolitiCap — belongs together.