Our Aggregation Approach
Weighted Moving Average
Each party's trend line is a recency- and sample-size-weighted average of every poll up to that date. Recency weight decays exponentially with a 25-day half-life (a poll from 25 days ago counts half as much as today's), and sample-size weight scales with the square root of (sample size ÷ 1,000) so a single huge poll can't swamp several smaller ones.
House-Effect Correction
Different pollsters consistently read some parties a point or two higher or lower than the field average — a "house effect." We estimate each firm's house effect as its average leave-one-out deviation from every other poll around the same date, shrunk toward zero for firms with few polls, and subtract it before averaging so the trend isn't just tracking which pollster happened to publish most recently.
Margin of Error
The shaded bands show the 95% margin of error of the weighted average at each point, not a single poll's own margin. Because real pollster disagreement runs wider than sampling error alone would predict, the underlying sampling-based uncertainty is inflated by a fixed design factor before the band is drawn — a deliberately conservative (wider) band rather than an optimistic one.
Data Smoothing
Cosine interpolation is applied between real polling dates to create smooth transitions on the daily chart, mimicking the S-curve of how public opinion actually shifts — slowly at first, then accelerating, then levelling off — rather than a straight line jumping between poll readings. The uncertainty band also widens the further a day sits from the nearest real poll.
Seat Projection Methodology
Our seat projections run a Monte Carlo simulation with 5,000 trials. Each trial draws a simulated vote share for every party with randomised polling error correlated across related parties (government vs opposition, established parties vs independents) — matching how real polling misses tend to move together — then converts vote share to seats using each party's historical vote-to-seat efficiency ratio, further adjusted for how favourable the current poll mix is for that party's PR-STV transfers, with wider uncertainty applied to parties with fewer historical elections behind their efficiency estimate. Results show the median outcome with 10th–90th percentile ranges.
Model Backtesting
The vote-to-seat conversion mechanism is checked against real history, not just asserted to work: for each of the last 5 general elections, the efficiency-factor lookup is recomputed using only the other 4 elections and applied to that election's real vote shares, then compared to the seats actually won. Averaged across every party in all 5 backtested elections, the model is off by ±3.2 seats per party — see the full year-by-year breakdown on the Seat Projections page. The uncertainty band itself is checked the same way: re-running the model's per-trial jitter on each election's real vote shares (no polling-swing noise, since that's not meaningful for a past result), the real seat count landed inside the model's own nominal 80% range 84% of the time and inside its 50% range 55% of the time — a stated "80% likely" range that's actually right about 80% of the time, not just a plausible-looking band.
Coalition Analysis
Coalition likelihood is assessed by summing each party's simulated seats across every trial and checking how often the combination reaches the 88-seat majority threshold. "High" likelihood means the coalition reaches a majority in most simulations, "Medium" means it does so occasionally, and "Low" means it rarely or never does.
Trend Calculations
Short-term trends compare the current weighted average against the average from a specified number of days ago (default 30 days). This provides a clear picture of momentum shifts.
Data Sources: Polling data is collected from Red C (Business Post), Ireland Thinks (Sunday Independent), and Ipsos B&A (The Irish Times). Historical election results from the official Irish election database. All data is publicly available.