# Taiwan Election Observatory — Methodology v0.1

## Purpose

The Observatory separates three questions: estimated current support, an election-day predictive distribution, and an explicitly conditional scenario. Taiwan's president is selected by nationwide plurality of valid votes; county wins are not used to determine the winner.

## Models

- **A — Polling Model:** combines only polls matching the selected candidate configuration and available by the cutoff. Demonstration weights reflect fieldwork recency and sample size. Candidate responses are normalised to valid-vote shares without claiming that undecided respondents split proportionally in reality.
- **B — Fundamentals Model:** the first release contains no fitted research coefficients. Demonstration mode uses a clearly synthetic prior. Research mode marks B unavailable until chronological election-cycle evaluation is possible.
- **C — Dynamic Event Model:** begins from the same latent support state observed by polls. Only approved events first public after the latest included poll can enter the baseline, reducing direct poll/event double counting. Effects use an identified log-share structure, decay through time, and remain uncalibrated assumptions.
- **X — Consolidated Forecast:** uses mixture sampling. One eligible component is selected according to nonnegative normalised weights, then a complete candidate-share outcome is drawn from that component. Interval endpoints are never averaged.

## Uncertainty

Current-support uncertainty includes sampling and additional shared polling error. Election-day forecasts add uncertainty for future movement. Sparse evidence does not produce narrow intervals. Demonstration output uses seeded Monte Carlo simulation and is mechanically reproducible.

## Data availability and vintages

Poll records preserve publication time, fieldwork dates, sample, mode, population, wording, candidate lineup, percentages, undecideds, weighting disclosure and source. Event records distinguish occurrence, first public availability, ingestion and revisions. Historical replay must exclude data unavailable at the selected cutoff.

## Event review

Event occurrence confidence is distinct from confidence in electoral effect. Categories organise evidence but are not treated as ten independently estimable factors. Related reports share a cluster ID. Real-world event effects require human review and are never assigned automatically.

## Limitations

The current public release is a working demonstration and research-data scaffold. It has not been historically calibrated or prospectively validated. No real future candidate list, official election date, live poll, economic observation or predictive-accuracy claim is supplied.