The forecast fits one statistical model to every published poll of the certified field and to seven past Toronto mayoral campaigns, then simulates the election thousands of times. The margin between the two poll leaders, each candidate’s vote range and each candidate’s chance of winning are three summaries of the same simulations. Toronto election history sets the uncertainty; it does not choose the winner.
Once the candidate field is certified, the model uses every published poll that asked about that field, each entered once. Polls taken before certification, with other names on the ballot, are left out; a check showed they would move the answer by about a point.
Each poll is read as a noisy measurement of where the race stands on its fieldwork dates, with an allowance for the firm that ran it and for how many people it asked. Three named candidates are modelled individually; the other certified candidates are one pool whose combined share is learned from past races.
Use every published poll of the certified candidate field, entered once each. Polls of earlier fields are not used.
Estimate each candidate's support as a path through the campaign, allowing for the pollster behind each survey and how many people it asked.
Seven past Toronto mayoral races set how much support moves week to week, how far pollsters sit apart, and how far final polls have missed the result.
Carry each candidate's path to election day and add the historical polling miss, producing sixteen thousand plausible full-ballot results.
Publish only when the model's numerical checks pass. Otherwise the previous forecast stays up rather than a broken one going out.
The margin between the two poll leaders, each candidate's vote range, and each candidate's chance of winning all come from the same simulated elections.
Seven past Toronto mayoral campaigns supply what the current polls cannot reveal by themselves: how much support typically moves from week to week, how far pollsters typically sit from one another, and how far final polls have landed from the eventual result. That last gap has averaged about sixteen points on the margin between the two leaders, and it is the largest reason a clear polling lead is not a near-certain win. The history calibrates uncertainty; it does not dictate who wins this election.
Before publishing, the model must pass its own numerical checks. If they fail, the previous forecast stays up. Alternative assumptions are refitted and kept as an audit record rather than shown as competing forecasts.
Because the three views summarize the same simulations, they always agree with one another. A candidate can finish ahead of one rival in many outcomes without winning the whole race in as many.
Each dot shows one candidate’s support among respondents naming a candidate, positioned at the poll’s fieldwork date. Decided and decided-and-leaning shares retain the poll’s published values. All-respondent readings are divided by the sum of every reported candidate choice, including other candidates, excluding undecided and non-voter responses. Tooltips identify derived shares and show the original percentages, which remain in the archive.
This conversion does not recover leaners the poll did not ask for. Question wording and offered candidates still differ. Polls without a known denominator or a complete response breakdown stay in the archive and are excluded from this chart. The LOESS curve follows the local shape of those dots instead of drawing a straight segment from one poll to the next.
The default “Since nominations closed” view shows polls reporting Chow, Bradford and Alexander with fieldwork completed after the August 21 nomination deadline. “All polls” includes the earlier history. Both views use the same LOESS curves fitted from the full comparable polling history; the default view shows only their portion since nominations closed.
The forecast-history chart uses the same toggle and defaults to updates published since nominations closed. Its dots use publication dates, when each poll could first affect the forecast, rather than fieldwork dates. Its two views share the full-history LOESS curves.
Each candidate is fitted independently. If a poll did not test a candidate, it contributes no dot and no inferred zero. Curves stop at the first and last observed dates, and a candidate with too few distinct observations appears as dots only.
The curve makes the polling record easier to read; it is not the election forecast.
See the polling record →Six reported poll values are shown as dots. A smooth line follows their local direction without passing through every dot.
The site puts open seats first, then compares scoreable incumbents using three measurements from their latest win. Candidate history, ward facts, and any ward polls add context beside that comparison; they are never blended into a hidden prediction of who will win.
No incumbent is running, so the ward goes to the top of the attention order.
Compares vote share, eligible-elector support, and electorate growth against the winning margin.
Shows only candidate facts that match a separately supported historical association.
Preserves reported snapshots when they exist; they do not become a ward forecast.
The Councillor Defeatability Index was developed by Matt Elliott for City Hall Watcher. Each measurement is ranked against the other scoreable Toronto incumbents, then the ranks are added with equal weight. A higher total means more comparative exposure on those measurements, not a forecast of defeat.
These are facts from Dianne Saxe’s 2022 win and the current electorate estimate, not a claim about who will win in 2026.
Saxe’s share of valid council votes in the win.
Votes for Saxe as a share of everyone eligible to vote.
Estimated additional electors compared with her 123-vote winning margin.
The first two values are among the lowest for Toronto incumbents, while estimated electorate growth is far larger than the previous winning margin. Together they explain the high index result and the separate growth flag without counting the same fact twice.
Candidate-history hints appear only when a confirmed fact matches a separately supported historical association. They describe context; they are not causal claims and do not enter the Councillor Defeatability Index.
Ward polls are preserved as reported snapshots when available. The site does not combine them with the incumbent index or candidate history to create a ward-level win probability.
An absent number or label is not zero. It means the evidence is too thin, a required calculation cannot run, or a candidate identity cannot be confirmed. Small differences between model checks do not make the estimate disappear.
If the current polling evidence is too thin, the forecast remains unavailable. Missing required calculations also withhold a quantity; we assess differences between completed checks behind the scenes. Public polls can still miss late movement, turnout differences, or a systematic error shared across firms, and the current election can behave differently from the past.
The primary comparison uses candidates with confirmed identities in the 2010, 2014, and 2022 stable-boundary general elections. It accounts for election, the candidate’s role in the race, and field size, and groups uncertainty by council contest.
A categorical idea needs at least five matching candidacies in five contests and two elections, plus an adequate comparison group. A numeric idea needs at least 20 observations in ten contests and all three elections. Coverage only opens the door: the plausible effect must also stay entirely on one side of no relationship.
The two Returning-councillor findings in open contests use an explicitly labelled small-sample tier: four contests, the same direction in all three elections, and additional influence and permutation checks.
Definitions that count every race except council, or add council only for some candidates, are not eligible for the public catalog. Aggregate history must match the complete confirmed history a reader can see.
Define the candidate fact, comparison group, race type, and direction before looking at the result.
Require enough candidates, contests, and elections, with confirmed links to their earlier races.
Compare vote share after accounting for election, candidate role, and field size, with uncertainty grouped by contest.
Record whether the direction repeats across 2010, 2014, and 2022 or appears only in the pooled result.
Publish only a standalone statement the evidence can carry; retain every other tested definition in the audit.
17 candidacies · 16 contests · 3 elections
Associated with more council vote share in the same direction in 2010, 2014, and 2022.
13 candidacies · 12 contests · 3 elections
For non-incumbent, non-returning candidates with prior races, having at least one win was associated with more council vote share in all three elections.
107 candidacies · 68 contests · 3 elections
Non-incumbent, non-returning candidates with any confirmed previous race received more council vote share than otherwise comparable first-time candidates in all three elections.
88 candidacies · 60 contests · 3 elections
The audit retains this as a pair—stronger than having no previous race, but weaker than other prior experience—but the ward cards do not turn that awkward comparison into candidate copy.
4 candidacies · 4 contests · 3 elections
The small sample was approved at an explicit limited-evidence tier after the direction repeated in all three elections and passed additional influence and permutation checks.
1 candidacy · 1 contest · 1 election
The estimated difference was large, but one example cannot support a public candidate-level statement.
199 candidacies · 110 contests · 3 elections
Across all candidate types, the uncertainty included no relationship. Among prior winners, additional wins showed no clear dose response.
16 candidacies · 14 contests · 3 elections
The uncertainty included no relationship, and the estimated direction did not repeat across all three elections.
“Withheld” does not mean disproven. It means the available coverage, identity evidence, or uncertainty cannot carry a public candidate-specific statement. The corrected reader-readable screen tested 21 definitions and approved all 10 that cleared its evidence gate. Together with the two previously approved findings, that produces the current 12-flag catalog.
Public polls can miss late movement, turnout differences, or a systematic error shared across firms.
Historical polling misses guide uncertainty, but the current election can still behave differently from the past.
Boundary changes and short observable career windows limit which elections can be compared on equal footing.
Incumbent defeats are rare, leaving too little evidence for honest ward-level win probabilities.
A past race is attached only when the person link is confirmed; unresolved histories are not guessed from names.
A historical pattern can help describe context without proving that the candidate fact caused later vote share.
Official election and candidate records provide the foundation. Public poll releases supply current readings, Toronto election history calibrates forecast uncertainty, and versioned analytical work supplies the council index and candidate-history audit.
Official City of Toronto election results and candidate-registration data provide council results, eligible electors, margins, and the current candidate slate.
Public releases from Liaison Strategies, Forum Research, Mainstreet Research, Pallas Data, Abacus Data, and Ipsos. The Polls page retains the firm and fieldwork date for every reading.
Previous Toronto mayoral elections provide the omitted-candidate and poll-to-election uncertainty used by the current forecast.
Matt Elliott’s City Hall Watcher methodology supplies the three-component incumbent comparison reconstructed and backtested for this site.
The upstream Council Defeatability Index project retains the tested hint catalog, evidence status, sample coverage, uncertainty, and identity audit. This page reviews contract 2.1.0.
Ward polls offer a rough snapshot, not reliable odds of winning.
The latest Forum polls have small samples and combine phone interviews with an online panel that does not randomly select voters. To judge how wrong they could be, we have just 6 ward races from 2022: one pollster, one election.
The chart shows how far the poll’s leader could be ahead of or behind their strongest rival. Left of zero means a rival is ahead; right means the leader is ahead. Taller bars contain more simulated scenarios.
Only candidates named in the poll are modelled. “Other” stays in the published results. The chart cannot tell us whether a candidate outside that group will win.
We compare each named candidate’s poll share with their share of votes cast for the same group at the actual election. Candidates outside that group are excluded from both sides of the comparison.
We fit two error models to those 6 comparisons: Dirichlet and logistic-normal. Both allow polling error to vary between wards, with uncertainty about its typical size. The chart combines equal numbers of simulations from each. That equal weighting is our judgment; the historical evidence is too limited to determine reliable weights.
We checked each ward by fitting the models to the other wards and predicting the omitted result. Large misses remained, especially in Parkdale–High Park. All 6 poll leaders stayed ahead among the matched candidates, so these results do not directly show how often a leader loses.
The historical errors combine sampling, survey bias and campaign changes. We do not reduce them because a newer poll has more respondents. Applying those errors across elections and survey methods is already an assumption. Today’s Forum polls weight by age and gender, but the number of decided or leaning respondents is smaller than the headline sample. Effective sample size is not published.
Changing the statistical distribution moves some range endpoints by up to 40 percentage points. Treat these scenarios as rough estimates.
