The formula & sentiment conversion

Daily survey trackers are reduced to a composite net sentiment score on a −100 to +100 scale. This composite sentiment is converted to an NDA national vote share by a linear transformation calibrated against historical general election results:

Seat Allocation: In First-Past-The-Post (FPTP) single-member constituencies, vote share translates non-linearly into seats according to Kendall's extended cube law:

Clamping & Ratios: Vote share outputs are clamped to a realistic 25%–60% range to prevent extreme polling swings from producing impossible totals. Opposition vote share is split between the INDIA alliance and regional parties in a fixed baseline ratio.

Conversion models, exponents and weights

Five complementary model paradigms run in parallel to evaluate seat distribution. Where these models agree or diverge provides a robust measure of structural model uncertainty.

Why each model parameter was set where it was

Model parameters are anchored to historical general election outcomes rather than arbitrary curve fitting:

Extended Cube Law Exponent k = 2.655

Fitted jointly across the 2019 and 2024 general election results to balance majoritarian seat bonuses against multi-party alliance splits.

Softmax Swing Transfer Temperature

Fitted to constrain winner-take-all seat amplification at wide polling margins, acting as a conservative floor on seat swings.

Opposition Split Ratio INDIA: 73.9% / Others: 26.1%

Anchored to the 2024 general election result ratio among non-NDA vote shares to maintain realistic third-party seat distributions.

Baseline selection & historical anchors

The model is anchored on the national vote shares and seat totals of the 2019 and 2024 general elections. Vote share sentiment equations are calibrated directly against the mean survey readings over the 60-day window preceding each election date.

Calibration Anchors: 2019 (NDA Vote Share: 45.0%, 353 seats) and 2024 (NDA Vote Share: 43.6%, 293 seats). Baseline constants are fixed and published in full so anyone can rebase or verify model projections independently.

Monte Carlo simulation framework

To evaluate polling error and sample variance, each daily update runs 20,000 Monte Carlo draws. The published 90% confidence envelope represents the p5 to p95 seat range for NDA, INDIA, and regional party blocs.

The random seed is fixed at 42 to ensure every simulation run is deterministic and perfectly reproducible.

Data cadence and staleness

Daily fieldwork updates are processed from CVoter's published tracking polls covering 41 individual indicator questions across 7 core categories.

What happens when polling data is missing or delayed

• The last known survey reading is carried forward. Values are never reset to a zero baseline, preventing artificial seat swings during field holidays.
• The dataset status is marked provisional if any core tracker feed is missing for more than 7 consecutive days.
• Simulation confidence bounds widen automatically when survey sample size drops below baseline thresholds.

Known limitations

Stated plainly, because a reader who finds these on their own trusts everything else less.

1. This is not a forecast of an unannounced election. It models present public opinion projected onto a hypothetical immediate poll. It carries no official status and does not replace Election Commission of India results.
2. National uniform swing model. There is no live constituency-level polling in the model. Calibration anchors on national vote shares and seat totals.
3. Alliance arithmetic is static. Seat-sharing deals, candidate selection, local party defections, and candidate-level popularity effects are invisible to national survey trackers.
4. Two anchor points. Every fitted constant rests on the 2019 and 2024 general election outcomes. That is the binding constraint on precision, regardless of daily survey volume.
5. Event impacts are associations, not causes. Attributing a seat move to a news event means public survey responses shifted during that window, which is consistent with the event having an effect but does not prove causation.
6. Confidence envelopes cover polling variance only. The 90% Monte Carlo confidence interval captures survey sampling noise, but does not capture vote-to-seat model conversion error.

Citation

Projection figures and model outputs shown on this page are published under CC BY 4.0, and may be cited as “MyDataLabs Lok Sabha Projection Engine — mydatalabs.in”. The underlying dataset is not distributed.