台灣選舉預測平台

The historical preset uses 2024 vote shares as a reference. The default lower-TPP preset is a sensitivity assumption, not a current poll or a future candidate lineup. Presidential entries are hypothetical party tickets. Baseline uncertainty is 1.5pp per entry; each hypothetical event adds the entered mean shift plus a random shock with the selected standard deviation, then shares are normalized. Events are independent assumptions, not 30/60/90 verified news records. Each event is sampled as high (2× volatility) or low (0.4×). Influence decays with its age before election day using the chosen half-life; events in the final 14 days receive the chosen timing multiplier. These coefficients are sensitivity assumptions, not empirically established effects. Zero mean shifts describe symmetric hypothetical shocks, not an assessment of actual political events.

For the Legislative Yuan, the vote result is party-list share. The seat scenario allocates 34 party-list seats among named parties reaching 5%, with largest remainders. The pooled “Other / independent” share is excluded from list allocation; it is not a single eligible party. The 73 constituency profiles use official 2024 candidate records, using the CEC 2024 candidate shares and 6pp local noise. Six Indigenous seats are held at the explicit 2024 reference allocation (2 DPP, 3 KMT, 1 independent). These profiles are not verified constituency forecasts; local nominations, alliances and Indigenous vote changes are not modeled.

The displayed average is exactly one third of each model’s mean vote share, expected seats and probabilities. Its 80% intervals come from the equal-weight pooled distributions, not averaged endpoints. The main result assigns each constituency to the party with the highest simulated win probability, selects the most frequent complete 34-seat list allocation, and includes the six fixed Indigenous seats. This yields exactly 113 whole seats. Combined-model decisions use pooled equal-weight model outcomes, rather than rounding the three seat totals separately. Fractional mathematical expectations remain in model details. Exact district ties prefer the historical winner then party order; tied list allocations use a deterministic lexical order. “Largest party” splits ties equally and does not imply a majority; majority requires at least 57 seats.

Historical reference dataset · 2024 seat breakdown. Hypothetical simulation results do not validate the model. Research mode remains unavailable until verified event effects and election-specific calibration exist.

Per-event shifts and baseline inputs above are assumptions, not fitted coefficients. Lower-TPP defaults are DPP 46 / KMT 44 / TPP 10 for presidential votes, and 38 / 38 / 12 / 3 / 9 for legislative party-list votes. These are sensitivity choices, not measured vote-intention estimates. If TPP does not contest the presidency, its assumed share is split using your entered transfer percentage; no transfer is inferred from party support.