Scan2029 — Methodology & Reasoning

1. What this system claims — and doesn't

Scan2029 estimates, for any given day, the probability that each tracked figure wins the 2029 Indonesian presidential election. The estimate is not a prediction of a fixed future; it is a snapshot of evidence (polls, news coverage) passed through an explicit model of how Indonesian elections actually get decided — nominations and coalition math first, popularity second.

Until coalitions crystallize (~2028), the honest reading of any number here is a trend with wide error bars. Indonesian history is unkind to early forecasts: Jokowi was a provincial mayor three years before winning the presidency, and Dedi Mulyadi was absent from national polling eighteen months before topping a July 2026 survey. The model encodes this humility directly — see the noise term (§5) and the "field" entrant (§6).

2. Tracked candidates

The roster is the top four by national polling plus two structural candidates who hold permanent seats despite little or no national polling, because their nomination path is credible: Pramono Anung (Governor of Jakarta, acceptable to both the Megawati and Jokowi wings of PDI-P) and Sherly Tjoanda (Governor of North Maluku, elected 2024 on an eight-party NasDem-led coalition, unaffiliated and courted by several parties — the type of outsider provincial executive Indonesian politics has repeatedly elevated). Ganjar Pranowo was reviewed and excluded. Everyone else — AHY, Ganjar, ministers, unknown 2028 entrants — competes in the simulation as an aggregate "field" entrant (§6), and a tripwire (§3) watches for breakout names that should be promoted onto the board.

3. Data pipeline

4. The poll aggregate

Each candidate's headline "poll aggregate" is a weighted average of all national capres (first-choice) survey results:

• Recency decay: a 75-day half-life. A five-month-old poll carries ~25% of a fresh poll's weight. This is aggressive by design — the 2026 landscape moved fast (Prabowo's approval fell 30 points in nine months).
• Pollster quality weights: 1.2 (Litbang Kompas) down to 0.7 (default for minor or unrated firms), reflecting track record and transparency. Unverifiable or disputed surveys are excluded entirely.
• Momentum is the gap between the recent window (last 120 days) and the older window (120–300 days). It adjusts the score by at most ±12%, and is dampened by data density: a single recent poll gets one-third of the full effect (min(1, n_recent/3)), so one survey — however credible — cannot swing the forecast alone.
• News sentiment (30-day average of LLM-tagged coverage) adjusts by at most ±8%, dampened below 20 mentions. It is deliberately capped: LLM sentiment shows positivity bias, so it nudges rather than drives.

5. Monte Carlo simulation

The score from §4 measures popularity today. Winning in 2029 requires surviving nomination politics and thirty more months of events. The model handles this with a 10,000-run Monte Carlo:

1. Each run samples every scenario (§6) true/false by its probability, multiplying candidate scores by the scenario's effects.
2. Each candidate's score is then perturbed by lognormal noise (σ = 0.55) — wide enough that a candidate at 8% today still occasionally reaches the mid-20s in simulation. This is the mathematical expression of "2.5 years is a long time": it reflects how far early polls historically sit from final outcomes.
3. The winner of each run is the highest adjusted score; win probability is the share of runs won. The p10–p90 band shown on each card is the spread of simulated vote shares.

6. Scenarios — where judgment enters, explicitly

This is the layer most forecasts hide and Indonesia most requires. Raw polling answers "who is popular"; scenarios answer "who is even on the ballot, with what machinery." Every scenario, its probability, and its effects are published on the dashboard and editable — they are analyst judgment calls, stated in the open so you can disagree with a specific number rather than a black box.

Current reasoning behind the key entries:

• Prabowo runs again — 75%. Raised from an initial 65% on a character argument: five presidential pursuits across twenty years (2004 Golkar convention, 2009 VP, 2014, 2019, 2024), a party founded as his personal vehicle, no succession culture. Biography says he does not voluntarily hand over the peak he waited a lifetime for — at 77, with collapsed approval, or otherwise.
• Dedi defects and finds a vehicle despite Prabowo running — 35%, and only evaluated in worlds where Prabowo runs. KDM is a two-year Gerindra transplant (24 years in Golkar before 2023), so defection is thinkable — but Prabowo controls the party, can pressure coalition partners, and can dangle the VP slot to contain him. If Prabowo retires instead, this scenario is moot: Gerindra succession is open and Dedi is its most electable heir.
• Jokowi machine backs Dedi — 35%. The volunteer/endorsement network that moved 20+ points in 2024 must land somewhere; Gibran's weakness makes Dedi its most plausible beneficiary.
• PDI-P nominates Pramono — 50%. The likeliest resolution of the party's Puan-vs-electability dilemma, per its bench: Pramono won Jakarta while the party lost everywhere else, threatens no one dynastically, and bridges the Megawati and Jokowi camps.
• Anies secures a ticket — 55%; opposition consolidates behind him — 30%, conditional on the ticket. The Constitutional Court's abolition of the 20% nomination threshold (Jan 2025) makes ballot access easier than 2024; hostile coalition math keeps it uncertain. His dormant 2024 base (~25% of the national vote) is the upside tail.
• Economic discontent persists into 2028 — 50%. The perceived-prosperity gap driving Prabowo's approval collapse; drags incumbents, boosts outsider governors.

Scenarios support dependencies: a scenario can require another (consolidation behind Anies cannot fire without his ticket) or be gated by one (Dedi's vehicle problem only exists if Prabowo runs). Impossible combinations are excluded from every simulation.

The "field" entrant carries a fixed score of 12 — roughly the current combined polling of AHY, Ganjar, and rising ministers — and competes in every run. Its win share (~8–10%) is the model saying: this far out, the eventual winner may not be on the board yet.

7. What would change the numbers most

• Any national poll confirming or reversing the July 2026 SMRC finding that Dedi overtook Prabowo (the recent window is still thin).
• A signal about Prabowo's 2029 intentions — the single largest variable in the simulation.
• Coalition movement: Golkar's posture toward KDM, where the Jokowi network lands, PDI-P's internal resolution.
• Rulings on election law (nomination rules, threshold litigation).

8. Audit history — errors found and fixed

Forecast credibility comes from showing corrections, not hiding them.

Known open limitations: Pramono's aggregate rests on one minor-pollster reading while other surveys place him near zero (his probability is likely optimistic); LLM sentiment carries positivity bias (mitigated by the ±8% cap); poll question methodologies are mixed (§4).

9. Stack

Cloudflare Worker + D1 (SQLite), scheduled crons (daily 05:00 WIB pipeline; Monday poll sweep and syntheses), OpenAI GPT-4o-mini/GPT-4o for reading Indonesian media at scale, and a dependency-free hand-rolled Monte Carlo. All model math is deterministic, versioned code; all judgment (scenario probabilities, pollster weights) is data, published on the dashboard and editable without redeploying.