On election night, 28 November 2026, this page follows the count live. Until then it shows the 2026 races as they stand before the count, and replays real past nights from the CEC's own records with the model run as if live: pick a year, press play, open any race. Nothing in a replay is simulated, and nothing from later in the night leaks in.
2026 年 11 月 28 日開票當晚,本頁即時追蹤開票。在那之前,這裡顯示 2026 年各場選舉開票前的狀態,並以中選會原始紀錄重播真實的過去開票夜,模型如同現場即時運算:選一個年份、按下播放、點開任何一場選舉。重播沒有模擬數據,也不會用到之後才開出的結果。
2026 年 11 月 28 日開票彼暗,這頁會即時綴開票。佇彼進前,遮顯示 2026 年逐場選舉開票進前的狀態,閣用中選會的原始紀錄重播真正的過去開票暗,模型親像現場咧算:揀一个年份、揤播放、拍開任何一場選舉。重播無模擬的數字,嘛袂用著較晏才開出來的結果。
The dial is the chance each of a race's two main contenders wins, from the votes counted so far. Everyone else's votes are set aside. It starts wide and narrows as stations report. Decided means the lead is larger than every remaining elector could overturn, with no assumption about turnout.
How it is worked out. The stations already counted are compared with how the same villages voted before (for mayors, the previous mayoral and presidential races, averaged; for president, the previous presidential race). That swing is applied to the stations still out, with room for districts to move differently. See the About page for how well it did.
Top panel. Mayoral nights: one block per race, colored by the projected leader and shaded by how sure the model is (as on the map), with counts by party and how many each won last time; hover a block for the race, click it to open it below. Presidential nights: the national race itself, with its chance and margin charts (and, for 2024, the count watch).
Three-way races (2024: eight counties where 柯文哲 was close): the dial shows the two candidates likeliest to win, and the third candidate's chance is given beside it. The chances come from the projected shares and their uncertainty, and start at the 5th station counted. Count watch (2024): per candidate, the floor (highest count among the CEC and the reliable channels), the other channels' range, and the model's projected final with its 80% range.
Map. Before a race has any counted stations, its county is hatched in the party color of the incumbent. Once counting starts, it takes the color of the projected leader's own party (gray for independents, teal for 台灣民眾黨), shaded from light (toss-up) to solid (decided). 金門, 連江 and 澎湖 are drawn in insets.
Turnout view. "Show: Turnout" switches the cards, map and race panel to the projected final turnout, worked out the same way as the vote swing. On the 2018 and 2022 nights it was within about 2 points by the time 5–10% of stations were in. The "Ballots counted" line uses it to estimate what share of all ballots has been counted, a better progress measure than stations, since stations differ in size.
Council view (2018, 2022). Who would hold each council seat if counting stopped now. Each district elects several members; the candidates with the most votes win. A district electing 4 reserves a seat for a woman, plus one per 4 more seats (地方制度法 §33); "women's seat" marks a woman elected this way, "displaced" the man who loses his place. A seat is decided once even the most damaging split of the votes still out (every remaining elector voting, all for the closest challengers) couldn't change who holds it. Nothing is projected here.
Needle motion. The big needles move like a gauge: when the estimate changes they swing to it, and for a few seconds they wander across the chances that go with the middle half of the projected margin, then settle on the estimate. The wandering shows how uncertain the estimate is; it isn't new data. Off if your device asks for reduced motion.
Legislature view (2020, 2024). The 73 districts each get a needle, from the swing since the previous presidential and legislative races (averaged). The 34 party-list seats are shown as the most likely splits with their chances (parties at 5% or more, largest remainder). The six indigenous seats are shown like a council district. District needles are less sure-footed than the mayoral ones: districts are small and swing on candidates.
Feed. The model uses polling-station results whenever the CEC's live feed has them. If the feed only gives district (行政區) totals, it falls back to those: the same model, noticeably less certain early on, and a box at the top says so. The replays all use station results.
Caveats. The two contenders are taken from the final result here; for 2026 they are set in advance. Times are when the CEC logged each station's report (all contests at a station at once), which can trail what TV showed. 2018 counting ran past midnight in many places because referendums were counted the same night.
指針是每場選舉兩位主要候選人各自勝選的機率,依目前已開出的票計算,其他候選人的票先不計。一開始範圍很寬,隨著投開票所回報逐漸收斂。確定表示領先票數已超過所有尚未開票的選舉人數,不需要任何投票率假設。
怎麼算的。把已開出的投開票所,和同一村里過去的投票結果比較(縣市長:上一屆縣市長與總統選舉的平均;總統:上一屆總統選舉),把這個「擺盪」套用到尚未開出的投開票所,並允許各區有不同的變化。模型過去的準確度請見說明頁。
上方面板。縣市長選舉夜:每場選舉一格,顏色為預估領先者的政黨,深淺代表模型的把握(與地圖相同),並列出各黨領先數與上屆當選數;滑鼠移到格子上可看該場選舉,點一下即在下方打開。總統選舉夜:全國選情本身,附勝率與差距走勢圖(2024 年另有開票數字監看)。
三強鼎立(2024 年柯文哲票數接近的八個縣市):指針顯示最可能勝選的兩人,第三人的勝率另外標示。勝率來自預估得票率及其不確定性,從第 5 個投開票所開出後才顯示。開票數字監看(2024):每位候選人的「下限」(中選會與可靠頻道中的最高票數)、其他頻道的範圍,以及模型預估的最終票數與 80% 區間。
地圖。尚未開出任何投開票所時,以斜線標示現任者的政黨顏色。開票後改為預估領先者本人的政黨顏色(無黨籍為灰色、台灣民眾黨為青綠色),由淺(難分高下)到深(確定)。金門、連江與澎湖畫在小框內。
投票率檢視。「顯示:投票率」會把卡片、地圖與選舉面板換成預估的最終投票率,算法與得票擺盪相同。在 2018 與 2022 年,開出 5–10% 投開票所時誤差約 2 個百分點以內。「已開票數」一行用它來估計全部選票已開出的比例;由於各投開票所大小不同,這比投開票所數更能反映進度。
議員檢視(2018、2022)。如果此刻停止開票,每個議員席次會由誰拿下。每個選區應選數名,得票最高者當選。應選 4 名的選區保障 1 席婦女名額,每增 4 名再增 1 席(地方制度法第 33 條);「婦女保障」標示依此當選的女性,「被排擠」標示因此落選的男性。即使剩餘票數以最不利的方式分配(所有未開票的選舉人都投票,且全投給最接近的挑戰者)也無法改變當選人時,該席為確定。這裡不做任何預測。
指針擺動。大指針像儀表一樣移動:估計改變時會擺向新值,接下來幾秒在預估差距中間一半範圍所對應的勝率之間游移,再停在估計值上。游移呈現的是估計的不確定性,不是新資料。裝置設定減少動態效果時不擺動。
立委檢視(2020、2024)。73 個區域選區各有一支指針,依上一屆總統及立委選舉(兩者平均)以來的擺盪推估。34 席不分區以最可能的席次分配及其機率呈現(得票 5% 以上的政黨,最大餘數法)。6 席原住民立委以議員選區的方式呈現。區域選區較小、受候選人影響較大,指針沒有縣市長的穩。
資料。中選會即時資料有投開票所層級結果時,模型就用逐所結果;若只有鄉鎮市區合計,就改用合計:同一個模型,但早期明顯較不確定,頁面上方會有提示。所有重播都使用逐所結果。
限制。這裡的兩位主要候選人取自最終結果;2026 年會事先設定。時間是中選會登錄各投開票所回報的時間(同一所的所有選舉一起登錄),可能比電視畫面晚。2018 年因同晚開公投票,許多地方開票到午夜以後。
指針是逐場選舉兩位主要候選人個別當選的機率,照這馬已經開出來的票算,其他候選人的票先無算。起頭的範圍真闊,投開票所愈回報就愈收縮。確定的意思是贏的票數已經超過所有猶未開票的選舉人,免任何投票率的假設。
按怎算的。共已經開出來的投開票所,佮仝一个村里過去的投票結果比(縣市長:頂一屆縣市長佮總統選舉的平均;總統:頂一屆總統選舉),共這个「擺盪」套用到猶未開出來的投開票所,嘛予各區有無仝的變化。模型過去準無準,請看說明頁。
頂懸的面板。縣市長選舉暗:一場選舉一格,色是預估領先者的政黨,深淺代表模型的把握(佮地圖仝款),閣列出各黨領先的數佮頂屆當選的數;滑鼠囥佇格仔頂懸看會著彼場選舉,揤一下就佇下跤拍開。總統選舉暗:全國的選情本身,有贏面佮差距的走勢圖(2024 年閣有開票數字監看)。
三強鼎立(2024 年柯文哲票數相倚的八个縣市):指針顯示上有可能當選的兩人,第三人的贏面另外標。贏面是對預估得票率佮伊的無確定性來的,愛等第 5 个投開票所開出來才顯示。開票數字監看(2024):逐位候選人的「下限」(中選會佮可靠電視台內底上懸的票數)、其他電視台的範圍,佮模型預估上尾的票數佮 80% 區間。
地圖。猶未開出任何投開票所的時,用斜線標現任者的政黨色。開票了後改做預估領先者本人的政黨色(無黨籍是灰色、台灣民眾黨是青綠色),對淺(五五波)到深(確定)。金門、連江佮澎湖畫佇細框仔內。
投票率檢視。「顯示:投票率」會共卡片、地圖佮選舉面板換做預估上尾的投票率,算法佮得票擺盪仝款。2018 佮 2022 年,開出 5–10% 投開票所的時,誤差大約 2 个百分點以內。「已開票數」這逝用伊來估全部選票已經開出來偌濟;因為逐个投開票所大細無仝,這比投開票所的數較會反映進度。
議員檢視(2018、2022)。若這馬停止開票,逐席議員會予啥人提去。逐个選區應選幾若名,得票上懸的當選。應選 4 名的選區保障 1 席婦女名額,逐加 4 名閣加 1 席(地方制度法第 33 條);「婦女保障」標按呢當選的查某人,「予人擠掉」標因按呢落選的查埔人。就算賰的票用上僫的方式分(所有猶未開票的選舉人攏去投,閣攏投予上倚的挑戰者)嘛改變袂著當選人的時,彼席就是確定。遮無做任何預測。
立委檢視(2020、2024)。73 个區域選區逐个有一支指針,照頂一屆總統佮立委選舉(兩个平均)以來的擺盪推算。34 席不分區用上有可能的席次分配佮伊的機率呈現(得票 5% 以上的政黨,最大餘數法)。6 席原住民立委用議員選區的方式呈現。區域選區較細、受候選人影響較大,指針無縣市長的遐爾穩。
指針搖擺。大支指針親像儀表咧振動:估計改變的時會搖向新的值,紲落來幾秒鐘佇預估差距中間一半範圍對應的贏面之間游來游去,才停佇估計值頂懸。游來游去是咧表現估計的無確定性,毋是新的資料。裝置有設定減少動態效果的時袂搖。
資料。中選會的即時資料若有投開票所層級的結果,模型就用逐所的結果;若干焦有鄉鎮市區的合計,就改用合計:仝一个模型,毋過早期明顯較無確定,頁面頂懸會有提示。所有的重播攏用逐所的結果。
限制。遮的兩位主要候選人是對上尾的結果揀的;2026 年會事先設定。時間是中選會登錄逐个投開票所回報的時間(仝一所的所有選舉做伙登錄),可能比電視畫面較晏。2018 年因為仝暗開公投票,真濟所在開票開到半暝以後。