Clinton Has 85 Percent Chance of Winning

Summaries Written by FARAgent (AI) on February 16, 2026 · Pending Verification

In November 2016, millions of voters, donors, and campaign professionals were told that Hillary Clinton had the race in hand, 85 percent at the New York Times, about 70 percent at FiveThirtyEight, as high as 99 percent at Princeton. When Donald Trump won anyway, the damage was not physical but it was real: public trust in expert authority took a hit, Clinton allies argued that polling-fed confidence contributed to weak attention to Michigan and Wisconsin, and the press looked foolish for treating probability as prophecy. The assumption had seemed respectable for years. Polling averages had often beaten pundits, academic regression models looked cleaner than gut instinct, and forecasters could point to a decent record in earlier cycles.

The case for the assumption was never invented out of thin air. In 2016 the national polls were not wildly wrong, Clinton did win the popular vote by about two points, and several modelers, Silver among them, explicitly said Trump still had a meaningful chance. Forecasters argued that probabilities were being misread by the public and by journalists who translated "favored" into "certain." They also noted that election models had performed reasonably well over many contests, especially when aggregating many surveys instead of trusting any single poll. From that view, 2016 showed the limits of communication and the inevitability of uncertainty, not the collapse of statistical forecasting.

But the evidence against high-confidence election prediction has grown. State polls in the Upper Midwest missed in the same direction, low response rates raised doubts about whether samples really captured the electorate, and later research examined social desirability bias, education weighting problems, and late shifts among undecided voters. Critics inside and outside social science tied the miss to broader replication and measurement troubles in the field, arguing that elegant models can sit on shaky data. A substantial body of experts now treats claims of precise electoral odds with more suspicion, while defenders still maintain that probabilistic models remain useful if read as rough estimates rather than crystal balls.