by Stephen Fisher and Alan Renwick.

Our forecast has taken a dramatic turn. Last week our polling average had Remain at 51% after setting aside Don’t Knows. It has this week dropped a further two points to 49%. This means Leave is ahead in our polling average for the first time, with 51%.

The forecast share of the vote for Remain has correspondingly dropped from 53% to slightly over 50%.

The 95% prediction interval is still ±12 points. So we are now forecasting that both Leave and Remain will win between 38% and 62% of the vote.

The probability that Remain will win the referendum has fallen from 68% last week to just 51% this week.

The methods behind our forecast

The method behind this forecast is based on the historical experience of referendum polls and referendum outcomes in the UK and on the EU elsewhere, as discussed here.

Our polling average is constructed by taking the most recent poll from each company within the last two weeks. If a company uses both phone and online modes then both the most recent phone poll and most recent online poll are used. This applies just to both ORB and ICM this week. The current average is based on the results of just six polls, of which two were conducted by phone and four online. All polls are adjusted to account for the tendency for phone polls to be more favourable to Remain. This is done by adding 1.5 to the Remain share for online polls and subtracting the same amount for phone polls.

The diminished magnitude of this adjustment, down from 2.2 last week, is partly due to our narrowing of the time period for our analysis of the mode difference. All the polls after the Purdah period began on 27th May have been from just four companies (ICM, Opinium, ORB and YouGov). Since the same is true of the polls that make up our polling average and because there have been recent changes in polling methodology (see here and here) it makes sense to re-estimate the mode effect using just the sixteen polls with fieldwork finishing after 27th May. After accounting for the trend towards Leave and random-effects for pollster-mode combinations, we find that on average support for Remain was 2.9 points higher in the telephone polls than it was in the online polls. The corresponding gap was fully 9 points when we estimated it for our first forecast three months ago.

So what is going on?

These latest model results demonstrate that the referendum outcome is now highly unpredictable. Indeed, there are currently two particular sources of uncertainty to add to the usual ones. The first is uncertainty about the polls. In regular elections, the pollsters can calibrate their findings by comparing past polling with past election outcomes – but there is no such comparator here. The difficulties the pollsters face are illustrated by the tweaks to their methods that they have made in the past few weeks. But these tweaks do not account for the changes in our polling average. Most of the methodological changes have had the effect of boosting rather than diminishing the reported Remain share. However, the picture could change again if they alter their methods again or if other pollsters (re)join the fray.

Second, our model includes two elements drawn from the patterns observed in referendums around the democratic world in recent decades: first, opinion tends, on average, to shift slightly towards the status quo as voting day approaches; second, even the final polls tend slightly to underestimate the vote for the status quo. The pattern we have observed in the last few weeks – a gradual shift towards the change option of leaving the EU – clearly violates the first of these general patterns. Whether that will continue is unknowable.

One possibility is that Leave campaigners have successfully switched the logic of the debate around, such that many voters now see greater uncertainty in what is, on paper, the status quo option than in the change option. The alleged likelihood of Turkey’s accession to the EU might be part of that. We have in the past referred to such shifts as ‘reversion point reversals’: situations where people see the change option as the best way to preserve what they like about the status quo. Our data from past referendums show that, where reversion point reversals occur, opinion does often buck the international trend by shifting away from the (on-paper) status quo.

On the other hand, another possibility is that, while voters have in recent weeks been steeling themselves to take a big step into a different future, caution will ultimately win out in the polling stations. The pattern in Scotland in 2014 was one of a gradual increase in support for independence until the very end of the campaign, when some shift back towards the status quo took place.

As things stand, we simply cannot know which pattern will ensue over the remaining nine days of the campaign.