{"id":696,"date":"2017-07-09T22:27:55","date_gmt":"2017-07-09T21:27:55","guid":{"rendered":"https:\/\/marriott-stats.com\/nigels-blog\/?p=696"},"modified":"2026-01-02T16:20:52","modified_gmt":"2026-01-02T16:20:52","slug":"uk-general-election-2017-7-review-of-my-predictions","status":"publish","type":"post","link":"https:\/\/marriott-stats.com\/nigels-blog\/uk-general-election-2017-7-review-of-my-predictions\/","title":{"rendered":"UK General Election 2017 #7 – Review of my predictions"},"content":{"rendered":"

Exactly one month ago, the UK woke up to the news they had elected a hung parliament for the second time in three elections.\u00a0 For many forecasters including myself, this came as a surprise as I had been predicting a Conservative majority\u00a0of 100<\/strong> seats<\/a>.\u00a0 In the event, the largest ever polling underestimate of the Labour vote <\/a>was enough to see the Conservatives lose their majority.<\/p>\n

At the beginning of my commentary on election night itself, I defined success for my forecasts as being how close the number of Conservative seats was to my forecast of 375<\/strong><\/a>.\u00a0 I also stated that if the number of seats was in the 340s<\/strong> I would consider this to be a prediction error.\u00a0 The final outcome was 317<\/strong> seats so clearly that is a major prediction error.<\/p>\n

**Update 20th Sept 2025 – This post was originally published 9th July 2017.\u00a0 It remains one of my most important posts but it did contain some graphics which were unreadable in some formats and some typos so I’ve edited this post to correct these issues.\u00a0 I’ve also updated the text format to my current style guide and taken the opportunity to add links to subsequent posts which reinforce particular points made in this article.\u00a0 Any new links are pointed out**<\/em><\/span><\/p>\n

\"\"<\/p>\n

NB: Some people count the Speaker as a Conservative giving the Tories 318 seats but I follow the House of Commons convention and count him as Other.<\/em><\/p>\n

So why was my forecast so far out?\u00a0 To answer, I am going to start by considering some fundamental principles of forecasting.\u00a0 A prediction consists of two parts:<\/p>\n

    \n
  1. A set of INPUTS<\/strong>\u00a0which are thought to be predictors<\/li>\n
  2. A MODEL<\/strong> which converts the inputs into a prediction<\/li>\n<\/ol>\n

    For your prediction to be right, you need to have the right inputs and the right model.\u00a0 If either is wrong, then your prediction will be wrong.\u00a0 If both are wrong, you can sometimes end up with the right prediction by accident (through one error cancelling the other error)\u00a0but normally your prediction error will be even worse.<\/p>\n

    **Update 20th Sept 2025 – the scenario of wrong inputs + wrong model = right prediction is exactly what happened to me with my 2024 UK General Election forecast!\u00a0 See GE2017 section here<\/a> for more details but the model error was a result of parameter error as explained below**<\/span><\/em><\/span><\/p>\n

    Prediction ERROR can come from\u00a04 sources either singly or combined:-<\/p>\n

      \n
    1. INPUT ERROR<\/strong> – The inputs you use in your model turn out to be incorrect.<\/li>\n
    2. MODEL ERROR<\/strong> – The model you have used to turn inputs into predictions is incorrect.<\/li>\n
    3. PARAMETER ERROR<\/strong> – The model you have used is the right one but the parameters in the model\u00a0are incorrect.<\/li>\n
    4. RESIDUAL ERROR<\/strong> – You have the right model & inputs but what\u00a0your\u00a0model does not take into account results in an error.<\/li>\n<\/ol>\n

      Let’s list the potential sources of error in my 2017 general election forecast.<\/p>\n

        \n
      1. Input Error<\/strong> <\/span>– I only ever had access to publicly available data throughout the election campaign, namely the polls, previous election results and demographic information from the Census.\u00a0 Of these 3 inputs, two were known in advance (previous results and demographics) but voting intentions were subject to potential polling error.<\/li>\n
      2. Model Error<\/strong><\/span> – My model was based on 3 main assumptions which were<\/a>; uniform national swing would not work as this was a realignment election, the Leave vote in 2016 would be the main driver of the Conservative and tactical voting against the Conservatives would take place.\u00a0 If any of these did not happen then I run the risk of having the wrong model.<\/li>\n
      3. Parameter Error<\/strong> <\/span>– In my model, there were two key parameters; how to account for the Leave vote in 2016 and how to reallocate the Green and UKIP votes in the seats where these parties had stood down.\u00a0 I used the opinion polls to help estimate these parameters.\u00a0 In addition, I had a parameter for an expected polling er<\/a>ror based on historical errors in the polls.<\/li>\n
      4. Residual Error<\/strong> <\/span>– No model is ever perfect so residual error will always be present in a prediction.\u00a0 The question is whether what you failed to take into account has a significant effect on the results.\u00a0 Two factors that emerged during the election that I did not take into account was the increase in voter registrations and suggestions that young people would be much more likely to vote this time than in previous elections.<\/li>\n<\/ol>\n

        It turned out that input error was by far and away the largest contributor to my prediction error followed by my assumption of a polling error in favour of the Conservatives.\u00a0 This was simple to detect with hindsight.\u00a0 As some readers of my election blogs may recall, one of the key inputs to my model was the regional vote swings as indicated by the polls<\/a>.\u00a0 \"\"What I did was to substitute the estimated swings in each region based on the polls with the actual swings seen on June 8th and then rerun my model.\u00a0 In effect, I was asking myself “what if the polls had got each region spot on, how much would my forecast have changed?<\/em>“.\u00a0 The outcome is shown in the table in the blue columns.<\/p>\n

        Two blue columns are shown.\u00a0 The first one is what my model would have predicted had the polls been bang on but I still included my expected polling error of a 2.6%<\/strong> underestimate in the CON<\/span>–<\/span>LAB<\/span><\/strong> lead<\/a>.\u00a0 The second blue column takes out my expected poll error.\u00a0 Straightaway, you can see that my predicted number of seats for the Conservatives would have been more or less bang on.\u00a0 Yes other parties are not so well forecast but my main criteria of success was the number of Conservative seats.<\/p>\n

        The polling error of 2017 is responsible for 56<\/strong> of the 58<\/strong> seat error I made for the Conservatives.\u00a0 I have already looked at this at the national level in a separate post<\/a> but it is worth looking at it again, this time\u00a0at a regional level.\u00a0 If we do\u00a0this, a striking picture emerges.<\/p>\n

        \"\"<\/p>\n

        I find this chart utterly galling.\u00a0 The basic premise of my prediction (and many other forecasters) is the Conservatives would hold or even lose a few seats in the South and make significant gains in the North & Midlands which voted heavily for Leave in 2016.\u00a0 The polls actually got the South and Scotland right (any error less than +\/-3%<\/strong> has to be marked as correct) but in the North, Midlands and Wales, the Labour vote was significantly underestimated.\u00a0 Labour was also underestimated in London but I am not overly bothered about this as London was an area where I was unsure if polls could fully capture the extent of potential tactical voting.<\/p>\n

        It is worth showing the predicted and actual swings by region to see where these errors came from.\u00a0 The top chart below is the actual swings, the bottom chart is the expected swings.<\/p>\n

        \"\"<\/p>\n

        In some ways, these charts further confirm I had the right model but the wrong inputs.\u00a0 In the South & Scotland where the polls were right, I predicted the Conservatives would make a net gain of 3 <\/strong><\/span>seats which was very close to the actual result of 0 <\/strong>net gains.\u00a0 In London, Wales, Midlands & North, I was predicting net gains of 42<\/strong><\/span> seats for the Conservatives and the final outcome was a net loss of 13<\/strong><\/span> seats.\u00a0 The above two charts explain why this happened very clearly.\u00a0 In the Midlands and North, the polls showed the Conservatives making larger gains than Labour which would deliver many gains.\u00a0 In reality, Labour managed to match or exceed the Conservative gains and thus denied them their expected gains.\u00a0 Why this has happened is something I will be investigating in more detail in a later post but as I say, I find it very galling the polls should get it wrong in precisely the areas where I expected the Conservatives to make their gains.<\/p>\n

        **Update 20th Sept 2025 – In this X\/Twitter thread in 2019<\/a>, I showed the Conservatives did not gain from UKIP standing down their candidate in some seats.\u00a0 UKIP15 voters who were left without the option to vote for a UKIP17 candidate split their vote 50:50 between CON17 & LAB17.**<\/em><\/p>\n

        I can’t claim my model was perfect but it was clearly good enough to capture the essential dynamic of the 2017 election and would have been right but for better polling data.\u00a0 As an independent statistician I have no access to any data other than what is in the public domain.\u00a0 The best I can do is try to estimate if the polls will be right or wrong and if so, how?\u00a0 In 2015, I correctly said the polls would underestimate the Conservatives<\/a> and I expected the same to happen to a lesser extent in 2017.\u00a0 In fact the opposite happened but at the beginning of May, I did state that there was a 25% chance of a polling\u00a0error favouring Labour<\/a>.\u00a0 I even went as far as saying that I would test a scenario assuming a polling\u00a0error favouring Labour similar to that seen in 1983 (which I did do but in a different way than I expected<\/a>) but what actually happened still exceeded that and did so in the worst possible way on a regional basis.<\/p>\n

        To make myself feel better, I am going to finish off by listing what I did get right!<\/p>\n