Methodology
How we model elections, produce forecasts and design strategies and marketing campaigns.
2. Electoral systems & seat allocation
3. Poll aggregation
4. Fundamentals
5. Social segments
6. Events & timing
7. Ensemble & Monte Carlo
8. Markov voter flows
9. Game theory
10. Dynamic programming
11. Strategy & marketing method
12. Validation & limits
1. Data layer
For every country in Europe, North, Central and South America and the major electoral democracies of Africa we store the head-of-state and legislative systems, regional executives and assemblies, municipal councils and mayors, and European Parliament seats — each with the seat formula, threshold, term, bonus or run-off rule and the years it was in force. Elections are stored at national, regional and municipal level with turnout and party results, so that each historical contest is analysed under the rule that applied at the time. The target depth is 50+ years.
Each country and region carries a profile: population and registered voters, age structure, urbanisation, migration, GDP, growth, unemployment, inflation, inequality, education, religion and language composition, union density, internet and social-media penetration, press freedom, trust, democracy and corruption indices; a narrative of history, society, culture, economy, education, anthropology and political culture; the main cleavages and issues; the main media with reach, ownership and lean; social-platform penetration; and external influences from alliances and international organisations (EU, NATO, IMF, AU, OAS, regional blocs). Sources include International IDEA, IFES ElectionGuide, ParlGov, CLEA, IPU Parline, World Bank, Eurostat, UN DESA, V-Dem, RSF, Transparency International, Reuters Institute and regional barometers. Seeded rows are marked to verify until an analyst confirms them.
2. Electoral systems & seat allocation
Highest-average methods allocate seats one by one to the party with the largest quotient v / d(s), where s is the number of seats already won:
Largest-remainder methods use a quota q (Hare V/S, Droop ⌊V/(S+1)⌋+1), give each party ⌊v/q⌋ seats and the rest by largest remainders. Legal thresholds are applied first; the “Others” bucket never receives seats.
Majority-bonus systems are modelled explicitly. Example — Greece (Law 4804/2021): the first party with at least 25% receives 20 bonus seats plus one seat per 0.5 points above 25%, up to 50 seats at 40%; the remaining seats are shared by Hare largest remainder among parties above 3%. Applied to the published June 2023 vote shares, the engine reproduces the actual result exactly (ND 158, SYRIZA 47, PASOK 32, KKE 21, Spartiates 12, Elliniki Lysi 12, Niki 10, Plefsi Eleftherias 8). Italian and French municipal bonuses and San Marino’s majority rule are implemented similarly.
For single-member plurality and two-round systems without district-level inputs we use the generalised cube law seats ∝ vᵏ (k = 3 for FPTP, 2.5 for two-round), and parallel systems combine both tiers. When constituency results are loaded, district-level simulation replaces this approximation. We also report the effective number of parties 1/Σp² and the Gallagher index √(½Σ(vᵢ−sᵢ)²).
3. Poll aggregation
Each poll’s shares are normalised to decided voters and weighted by recency, sample size and pollster quality:
With three or more pollsters, house effects are estimated iteratively as each pollster’s weighted deviation from the consensus, shrunk by k/(k+2) for pollsters with k polls, and removed before re-averaging. The standard error combines sampling variance Σwᵢ²·p(1−p)/nᵢ / (Σwᵢ)² and between-poll dispersion divided by the effective number of polls (Σw)²/Σw². Internal campaign polls stay private to their campaign.
4. Fundamentals
A transparent prior for each party starts from its previous result and adjusts for leader popularity and, for the governing party, the economy and the cost of ruling. Default coefficients (calibrate per country on historical data):
| Factor | Effect |
|---|---|
| Leader net approval | +0.08 pp per point (≈ +0.8 pp per +10) |
| GDP growth (incumbent) | +0.6 pp per point |
| Change in unemployment (incumbent) | −0.9 pp per point |
| Inflation above 2% (incumbent) | −0.35 pp per point |
| Years in office beyond 4 (incumbent) | −0.5 pp per year |
5. Social segments (class, education, age, urban/rural …)
Where cross-tabulated support is available, a post-stratification model (a simplified form of MRP) combines segment support with population weights from the indicator layer and segment turnout:
It also reports turnout sensitivity — how each party’s share moves if a segment’s turnout rises by 10% — which feeds mobilisation strategy.
6. Events, social reactions and timing
Scandals, debates, protests, endorsements, economic shocks or international crises are logged with an estimated impact in percentage points and a half-life; their effect decays as impact·0.5^(days/half-life). Timing enters through the time-to-election weight below and through the dynamic-programming tools in §10.
7. Ensemble and Monte Carlo simulation
Components are blended with weights that shift toward polls as the election approaches:
Uncertainty for each party combines the poll standard error with a systematic polling error that grows with time to election: σ = √(SE² + (e₀·√(μ/25))²·(1 + d/100)), with e₀ = 2 pp by default. We draw 10,000 scenarios including a small correlated industry-wide error, normalise each draw, and pass it through the country’s seat rule. Outputs: mean and 90% interval of vote share and seats, probability of first place, of crossing the threshold, of an outright majority, and of no single-party majority.
8. Markov voter flows
Voters move between parties and abstention according to a row-stochastic matrix P estimated from panel surveys, ecological inference or expert judgement (or built from retention rates and party attractiveness). The distribution evolves as vₜ₊₁ = vₜ·P; we report the trajectory, the stationary distribution, mixing time and the largest net flows — the pools a campaign can realistically win from or lose to.
9. Game theory
Positioning and message games. Choices such as which issue to lead on are modelled as a payoff matrix against an opponent’s choices. Zero-sum games are solved by fictitious play, which converges to the optimal mixed strategies and returns the game value with bounds; non-zero-sum games return pure Nash equilibria and the mixed equilibrium of 2×2 games.
Resource competition (Colonel Blotto). Two campaigns split resource units across regions. Region r is won with probability σ(leanᵣ + εᵣ·(ln(1+xᵣ) − ln(1+yᵣ))); payoff is expected seats. Each side’s best response is computed exactly by dynamic programming, and iterated best responses against the opponent’s average converge to equilibrium allocations.
10. Dynamic programming
Budget allocation. Each channel or region has diminishing returns g(x) = cap·(1 − e^(−x/sat)). The optimal split of B units solves F(j,b) = maxₓ F(j−1,b−x) + gⱼ(x), with minimum and maximum constraints; we report marginal returns and compare with an equal split.
Spend scheduling. Persuasion decays toward election day by δ per week: Vₜ(b) = maxₓ [g(x)·δ^(T−t) + Vₜ₊₁(b−x)].
Election timing. For a government choosing when to call an election, popularity follows a Markov chain and each state has a win probability; optimal stopping Vₜ(s) = max(win(s), Σ P(s,s′)·Vₜ₊₁(s′)) yields a threshold policy for each period.
11. Strategy and marketing-campaign method
- Situation analysis — the electoral rule and its incentives, historical results and swing regions, country/region profile, cleavages, media landscape, external influences.
- Segmentation — aggregate segments (geography, age, education, class, urban/rural) with size, turnout and support; never individual profiling.
- Positioning and key messages — informed by the message game and segment sensitivities.
- Objectives — measurable (vote share, seats, turnout, awareness, favourability, reach, volunteers, donations) with baseline, target, weight, priority and due date; weighted progress is tracked automatically.
- Actions and channels — TV, radio, print, social, search, digital ads, door-to-door, phone, rallies, debates, press; budget per action; optimal allocation and timing from §10; regional effort from §9.
- Marketing plan — channel mix, aggregate audiences and creative brief, compliant with political-advertising transparency rules.
- Procedures — legal checklist for filings, ballots, observers, finance reporting, counting and challenges, with deadlines.
- Feedback loop — new polls, events and media monitoring update the forecast; objectives are re-scored and resources re-optimised.
12. Validation, limitations and ethics
Seat-allocation code is tested against official results (e.g. Greece, June 2023). Forecast coefficients should be back-tested per country on past elections before operational use. Model outputs are only as good as their inputs: unrepresentative polls, missing regional data or mis-specified matrices produce misleading results, which is why every output shows its inputs and uncertainty. The Platform models aggregates, not individuals, and its Terms prohibit disinformation, voter suppression and unlawful targeting.