While most exit polls forecast that Narendra Modi would become India's next prime minister, only one little-known outlier accurately predicted the scale of his overwhelming election victory.
The secret recipe of New Delhi-based Today's Chanakya group was careful interview selection, ensuring that its survey accurately represented the demography of any given constituency in the world's largest and most diverse democracy.
"In India, when any poll is released, everyone talks about the sample (size) numbers ... no one asks about the quality of the sample," V.K. Bajaj, chief executive officer at the polling group, told Reuters.
"Every state has some caste combinations prevailing ... we tried to make the sample as representative as we could."
With an electorate of 815 million people, India's election was phased over several weeks in different parts of the country, and pollsters released their exit polls after the final phase on the evening of May 12, and the results were announced on May 16.
Bajaj's polling group predicted the Modi-led coalition would win 340 of the 543 parliamentary seats, with an error margin of 14 seats, and 291 seats for the Hindu nationalist BJP alone.
The alliance won 336 of the 543 seats in India's lower house of parliament, with the BJP itself managing to cross the halfway mark with 282 seats.
All other post-poll analyses forecast that Modi's coalition would win fewer than 300 seats.
Exit polls have a patchy track record in India, exaggerating the BJP's seat share in the 2004 and 2009 elections. In both cases, the Congress party formed coalition governments.
Today's Chanakya, part of a family-run research firm started two decades ago, interviewed 38,984 voters. Some 250 field representatives worked with its 175-strong team to arrive at its prediction, Bajaj said.
None of the polls taken before and during voting had predicted the scale of the BJP victory. A poll conducted by Hansa Research for the NDTV news channel came closest, predicting BJP and its allies would win a narrow majority of 275 seats.