Accuracy Report
Accuracy Report
An honest review of our pre-poll AI predictions against ECI's actual results across all 234 Tamil Nadu Assembly constituencies.
The seat-shape was wrong because TVK converted vote-share to seats far more efficiently than any pre-poll model predicted.
Source: Election Commission of India ↗
Every AI model, polling agency and political tracker in India under-counted TVK by 80+ seats. The wave was a category-five blind spot for the entire forecasting industry.
All 8 published pre-poll surveys plus VoteMatrix's AI consensus, ranked by date.
| Source | DMK+ | NDA | TVK |
|---|---|---|---|
IANS-Matrize | 104–114 | 114–127 | 6–12 |
News18-Vote Vibe | 113–123 | 106–116 | 2–8 |
Agni News Agency | 180+ | 50–60 | 0–10 |
VoteMatrix ★ | 170–200 | 10–45 | 2–12 |
Lokpoll | 181–189 | 38–42 | 8–10 |
News18-Vote Vibe | 90–100 | 130–140 | 2–6 |
Spick Media | 88–94 | 112–120 | 4–5 |
Vikatan | 121 | 83 | 3 |
Thanthi TV | 98–103 | 91–101 | 1–2 |
ECI ACTUAL ✓ | 73 | 53 | 108 |
Source: Wikipedia (2026 Tamil Nadu Legislative Assembly election). VoteMatrix sat in the optimistic-DMK camp alongside Lokpoll and Agni News Agency. Both camps missed TVK.
10 marquee constituencies where we named the eventual winner with the correct party.
Where we identified a coherent regional read, the model held: every NDA call we made in the western belt landed. AIADMK + PMK + BJP picked up these as a near-uniform block, exactly as our regional model projected.
Direct seat-call match rate, broken out by alliance.
| Alliance | Precision | Recall |
|---|---|---|
| DMK+ | 69 / 197 = 35% | 69 / 73 = 95% |
| NDA | 12 / 35 = 34% | 12 / 53 = 23% |
| TVK | 2 / 2 = 100% | 2 / 108 ≈ 2% |
95% recall on DMK+ means: of the 73 seats DMK alliance actually won, our model flagged 69 as DMK strongholds. The under-extension on TVK is what dragged precision down across the board.
Where our calls landed across the actual outcome. Diagonal cells (green) are matches; off-diagonal cells are misses.
Single largest miss bucket: 87 seats we called DMK+ that TVK actually won.
In every other constituency, our model placed the eventual winner among the three named candidates. The full miss list:
*Vanur may be a name-spelling false negative ("Gowtham D" vs our "Gautham Dravidamani" — likely the same candidate).
The model anchored DMK+ incumbency correctly — 95% of seats DMK alliance actually won were on our DMK list. Welfare-scheme delivery, alliance arithmetic, and 2024 LS signals proved a reliable floor for incumbent strongholds.
Where we identified coherent geographic blocs (Kongu/Salem-Dharmapuri NDA, Delta DMK), the model held cleanly. 12 of 12 NDA seats in the western belt landed.
Defection bets (OPS-on-DMK), shift bets (Senthil Balaji to Coimbatore South), and personality-anchored seats (EPS, Velumani, KN Nehru, Udhayanidhi) all paid off.
Our scenarios noted "TVK crosses 15–20% statewide but FPTP limits conversion." This assumption broke. TVK's vote was geographically concentrated enough — particularly in Chennai, Madurai, and the Kongu textile belt — to convert at far higher seat-per-vote-share rates than any historical third-party precedent suggested.
Chennai #14–#33 (excluding Harbour and Chepauk) and Madurai #189–#195 swept to TVK. Our model treated these as DMK Clean Sweeps with 40%+ floors. We missed a generational urban shift that no published pre-poll captured either.
Stalin (Kolathur), PTR (Madurai Central), TRB Rajaa (Mannargudi), Anbil Mahesh (Thiruverumbur), Geetha Jeevan (Thoothukudi), Vanathi Srinivasan (Coimbatore N) — all called as Clean Sweeps, all lost.
Three explicit corrections going forward: (1) Vote-share-to-seat conversion curves for new third parties calibrated on real geographic concentration data, not historical FPTP averages. (2) Urban anti-incumbency signals weighted independently of welfare-scheme delivery metrics. (3) Higher floor-uncertainty bands for incumbent ministers in volatile constituencies — no Clean Sweep tier without 8+ point cushion.
If you're working on election intelligence, market research, predictive modeling, or data-driven storytelling — we'd love to hear from you. Drop us a brief and we'll get back within 48 hours.
Prefer email? Reach us directly at votematrix@alphabsolutions.com