ascent-agri · daily agricultural market intelligence

Robusta Coffee Monitor

A model-driven read on coffee markets and the Vietnamese crop: market regime from a hidden-Markov engine, rainfall anomalies in the Central Highlands robusta belt, and producer-currency pressure from the Brazilian real. Rebuilt daily; everything open source.

crisis · 1 sessions · exposure guide ×0.40

Coffee (benchmark)

277¢/lb

1 week

-11.6%

1 month

-22.9%

Highlands rainfall

+1.6σ vs seasonal norm

BRL / USD

-0.7% over the past month

Coffee futures closed at 277¢/lb — down 11.6% on the week and down 22.9% on the month. The regime model reads the market as crisis and has for 1 sessions. Rainfall in the Central Highlands is running above its seasonal norm (+1.6σ) — favorable moisture for the robusta belt. The crop is in its fruit filling window, and stage-weighted stress is low. The Brazilian real is little changed on the month (-0.7%) — a neutral currency backdrop. Looking ahead: the 14-day forecast calls for 161 mm of rain in the robusta belt against a 170 mm seasonal norm (-0.1σ) — projected crop stress: low (Open-Meteo forecast model, issued 2026-09-18; skill scored publicly as windows close).

Market regime

price data through 2026-09-18

Coffee futures price with regime shading — Coffee futures (ICE arabica KC=F — the compliant daily benchmark; the hand-built roll-adjusted robusta series joins the page as contract history is assembled). Shading is the regime engine's smoothed state: a 3-state Gaussian HMM over price, currency and weather features, with asymmetric hysteresis so defensive transitions trigger faster than upgrades.

The long view

full price history through 2026-09-18

Full coffee price history with 200-day average — The market most of the trade actually discusses: price against its 200-day average, full history. Green-coffee buyers hedge and reprice off moves like these, not intraday noise — this view exists because an importer's education team told us that's the horizon they teach.

Growing conditions — Central Highlands, Vietnam

weather data through 2026-09-12 · Buon Ma Thuot, Dak Lak (12.68 N, 108.04 E) · 3% of the last 30 days near-rainless

crop stage: fruit filling · stage-weighted stress: low (0.16) — peak monsoon: moisture normally ample, so a deficit is unusual but damaging

Rainfall anomaly and dry-spell intensity at Buon Ma Thuot — Vietnam grows most of the world's robusta, and most of that in the Central Highlands. Each series is compared only to its own trailing year, so the anomaly is seasonal-aware and fully causal — and the stress read is phenology-weighted: the same rainfall deficit matters far more during dry-season flowering (Jan–Mar), when a failed blossom cannot be recovered, than during harvest (Oct–Dec), when wetness, not drought, is the risk. Persistent dry anomalies here have preceded the major robusta squeezes.

The next two weeks

Open-Meteo forecast model · issued 2026-09-18 · window 2026-09-18 → 2026-10-01 · crop stage: fruit filling

161 mm expected vs 170 mm seasonal norm (-0.1σ) — projected stress: low

14-day rainfall forecast vs seasonal norm

The only forward-looking panel on this page — and therefore the one that gets scored. Every issued forecast is committed to an append-only ledger before the outcome is known and verified against realized rainfall once the window closes (see the track record below). Until enough windows close, treat this as an unverified model output, not a fact.

Growing conditions — Sul de Minas, Brazil

weather data through 2026-09-12 · Varginha, Minas Gerais (−21.55, −45.43) · the arabica side of the market

Rainfall anomaly and minimum temperature at Varginha, Sul de Minas — Brazil dominates arabica, and Sul de Minas is its largest belt. Two risks matter here: dry spells (the same anomaly logic as the Vietnam panel) and winter frost — the dashed line marks the frost-risk threshold on daily minimum temperature. Most of the specialty trade buys arabica, so this panel pairs with the benchmark price above.

Currency driver — the Brazilian real

FX data through 2026-09-18

BRL per USD — Brazil is the largest coffee producer; growers sell dollars and spend reais. A weakening real raises local-currency revenue per bag and historically adds selling pressure to world prices — one of the regime model's inputs.

The ledger — the model in public

Every weekday the pipeline writes down what the model believes — regime call and target exposure — before the outcome is known, and commits it to the repository. Entries are never edited or deleted; scoring uses only the prices recorded in the ledger itself, with a 1-day execution delay. If the model is wrong, this section says so forever.

56 entries · 54 scored days (2026-06-25 → 2026-09-18) · model +0.11% vs buy-and-hold -10.83% · mean exposure 0.04

Forecast verification

40 closed 14-day windows · forecast MAE 27.9 mm vs climatology 39.7 mm · skill +0.30 (beats the climatology baseline) · bias -5.8 mm · stress-band hit rate 88% · raw snapshots.

Inspect the raw ledger on GitHub

Data & API

Organizations and builders can consume this monitor programmatically — everything below is regenerated with each daily run, versioned, and free under CC BY 4.0 (attribution required). Derived series only: regime labels, risk multipliers and weather anomalies computed by this project.

api/latest.json — today's regime, posture, exposure guide, growing-condition anomalies, and the daily brief (stable schema, versioned).
api/history.csv — the full daily history of regime labels, risk multipliers and rainfall anomalies.
feed.xml — the daily brief as RSS.
api/changes.json — state changes only (regime flips, stress-band crossings) — poll this on your schedule instead of scraping the page.
alerts.xml — the same changes as an RSS feed: fires only when something moved (pairs well with any RSS-to-email bridge).
brief/latest.html — the week on one printable page (Cmd/Ctrl+P → PDF); Friday editions are archived by date.
the public ledger — the model's append-only, scoreable track record.

Using this in your shop or newsletter? Open an issue or email — custom growing regions can be added on request.

Research

The question behind this page — do growing-region weather anomalies actually predict coffee futures returns? — is tested properly in the project's working paper: an event study with a-priori definitions, permutation inference, and a built-in placebo structure (each region's weather vs the other region's crop). Headline: Brazilian dry spells precede arabica rallies with the right sign at every horizon (+5.1pp over 5 days, n=4, uncorrected p=0.04); the placebo is null; the Vietnam→robusta test is still data-limited — and the famous 2021 frost exposed a real flaw in threshold-based event definitions.

Read the paper (PDF) · markdown + reproduction commands

Methods, honestly

This monitor is the public face of an open research project: a single-instrument port of a multi-agent equity platform, rebuilt for agricultural markets.

Regime: 3-state Gaussian HMM on trailing price, BRL/USD and weather features; entropy filter, asymmetric hysteresis, and a rule-based crisis override.
Weather: Open-Meteo daily observations at Buon Ma Thuot, z-scored against their own trailing year (causal — no future data).
Everything above is descriptive. The project's own walk-forward backtest of a trading strategy on these signals is deliberately published with its unflattering result (negative walk-forward efficiency) — the point is evaluation machinery that makes lying to yourself hard.
Code, tests (120), and the full backtest notebook: github.com/ScottDongKhang/ascent-agri.