Jump to content
270soft Forum

Poll of Polls ELO Forecasting 2020


Recommended Posts

I developed a forecasting model based in ELO, a system to created by Arpad Elo to calculate relative force rakings. How can this be useful to predict elections? My theory is that you can predict elections with a good poll of polls if you evaluate different pollsters between themselves.
I made the rankings starting in 2004 and continuing to 2012, trying to predict 2016. Im making 2 models one for GOP and one for DEMS. I finished the GOP part analysis from 2004 to 2012. The objective was to see the 2004 order of pollsters by the best to the worst predicting G.W.Bush results. If we have 30 different pollsters, the first wins 1vs 1 against the 2nd and all below, the last loses with everyone, and the 4th wins against all below losing to the podium. After I was going to 2008 and follow this way. The final national result was that...

Player name Rating
ABCWash Post 1,337
American Res. Group 1,183
AP-Ipsos 836
ARG 934
Associated PressGfK 1,187
Ayres McHenry 752
Battleground 1,411
Bloomberg 658
CBS News 1,231
CNN 962
CookRT Strategies 678
   
Democracy Corps (D) 825
DiageoHotline 1,221
EsquireYahoo! News 1,044
F&MHearst-Argyle 1,202
FOX News 1,043
Gallup 712
GWBattleground 1,108
Harris 1,098
Hartford CourantUConn 875
HotlineFD 705
   
IBDTIPP 953
ICR 991
InsiderAdvantage 651
LA Times 1,007
Marist 1,030
Mason-Dixon 1,179
McLaughlin (R) 662
Monmouth 1,159
National Journal 950
NBCWSJ 1,319
   
Newsweek 684
NPR 1,157
Pew Research 1,431
POSGQR 886
PPP (D) 1,076
Quinnipiac 971
Rasmussen 1,117
Reason-RupePSRAI 922
Resurgent Republic (R) 843
ReutersZogby 1,032
   
Sacred Heart Univ 1,010
Time 757
Univ. of Connecticut 812
WashTimesJZ Analytics 1,253
Zogby 856

With those ranks I did put with linear regression in excel putting the best pollster in left and the worst in right and all the others in the middle. Pollsters without information had 1000 rating. My second theory is that the best pollsters define the race direction and the worst are outliers. So, I did that with the last poll from the 23 independent pollsters from RCP. 0 intersection defines the final result. Very good job in 2016 predicting 46.2% Trump. I will do for 2020, and dems also. I ask if there is any error, pollsters names are a big confusion and can be repeated sometimes. Any big error, i need to recalculate everything.

image.png.2044eb48f85c2dec93beb00a61741176.pngmodels one for GOP

 

Link to post
Share on other sites
38 minutes ago, ThePotatoWalrus said:

Epic

 

31 minutes ago, vcczar said:

Very interesting 

Thanks!

I finished 2020 GOP, exluded 2019 pollsters that didnt update results. GOP poll of polls gives 43.8% for Donald Trump. I can update everyday and add states.
image.thumb.png.960286cf4feda1beaabee66d23b4a6df.png
 

Link to post
Share on other sites
22 minutes ago, Entrecampos said:

 

Thanks!

I finished 2020 GOP, exluded 2019 pollsters that didnt update results. GOP poll of polls gives 43.8% for Donald Trump. I can update everyday and add states.
image.thumb.png.960286cf4feda1beaabee66d23b4a6df.png
 

Yeah, I'd like to see the updates and the forecast for both Trump and Biden. We can see if my map or your map ends up being more accurate. 

Link to post
Share on other sites
38 minutes ago, vcczar said:

Yeah, I'd like to see the updates and the forecast for both Trump and Biden. We can see if my map or your map ends up being more accurate. 

I need to change my map completely, when I add polls to historical tendencies model (will be a mix), and i could adjust historical tendencias to be much more similar to yours, but I want to discover possible flips. I think your forecast as it is, is better overall, but mine can detect weird tendencies that most models will not detect (Like the state where you live in 2016).

EDIT> Results for 2004 GOP with the ELO Forecast. 49.6% Bush. 1.1% away. Much Better than RCP mean value. Now I can go to DEMs.
image.thumb.png.81f71c233a6a2e951aaea37c1c95addd.png
 

Link to post
Share on other sites
23 minutes ago, Entrecampos said:

DEMS 2016 result isnt bad, 47.0%, but it isnt perfect, its ok. 1.2% error. RCP mean was 46.8%, 1.4% error. The model works, even if it isnt perfect, can be a good contribute for predictions.

Untitled.png

I repaired in an error here, 36 in left row needs to be 35, 37 too, and 37 from marist needs to be 40, la times 37 needs to be 41, my fault.

 

6 minutes ago, vcczar said:

@Entrecampos So what's your popular vote projection for Biden vs. Trump in 2020?

I would say in moments.

Link to post
Share on other sites
3 minutes ago, Entrecampos said:

Here is Biden poll of polls, so 48.2% and 43.8% Trump. @vcczarimage.thumb.png.5005b4bf9f5336021e162f84bdd45954.png

 

So basically, Biden is projected to lead by only 60% of what he's currently polling vs. Trump nationally. 

Link to post
Share on other sites
8 minutes ago, vcczar said:

So basically, Biden is projected to lead by only 60% of what he's currently polling vs. Trump nationally. 

The Psychohistorian and Prophet Hari Seldon has calculated and spoken! Question not! :P

Link to post
Share on other sites
12 minutes ago, Entrecampos said:

I will improve the model with state polls, all polls from all the 50 states +1 published since 2004 will suffer a comparison. Only after this i can use the model to predict states.

I think using any polls prior to 2012 will skew any accuracy for 2020. The whole American political zeitgeist has changed since the passage of Obamacare and the rise of Birthers and etc., which were led by Trump. 

Link to post
Share on other sites

Today Biden ups to 48.6% and Trump downs to 43.4%.
I made a big revamp from the scratch in historical model for states. Im only using it, but i will add state polls, doing a mix history/polls when i finish analysis. Utah results are weird because McMullin 2016. But I respect the model for now.image.thumb.png.34924ed3e112eaacf1321055c62f0ce7.png


 

 

Link to post
Share on other sites
25 minutes ago, Entrecampos said:

Today Biden ups to 48.6% and Trump downs to 43.4%.
I made a big revamp from the scratch in historical model for states. Im only using it, but i will add state polls, doing a mix history/polls when i finish analysis. Utah results are weird because McMullin 2016. But I respect the model for now.image.thumb.png.34924ed3e112eaacf1321055c62f0ce7.png


 

 

2016 had an unusual # of 3rd party votes that make any relations between 2012 and 2020 difficult. It might be easier to just look at margin between the two major parties sometimes. UT is going to be 50%+ Trump. The model needs to show that to have any sort of reliability. You are showing about 8% 3rd party votes, which is incredibly unlikely in this election. Neither the Green nor the Libertarians have anyone with name recognition this time around. They'll likely drop their support by half. Kanye West is mainly a write-in. You need something in your model that can comprehend a drop in 3rd party votes. 

  • Upvote 1
Link to post
Share on other sites

Im increasing the data for ELO database. Adding some states poll analysis (Alabama 2004 and 2008, Alaska 2004, Arizona 2004,  I expect to finish all states from 2004 to 2016 before the 2020 election), and changing the base national methodology. I was comparing different pollsters between themselves, from the best to the worst in X election, predicting reps and dems. Now Im giving pollsters victories and defeats, if a Pollster is more than 10% out of reality, they lose 10 *games* lets say. If they are 3% from the real result, they win 7 games and lose 3, if they 5%, they win 5 and lose 5. And now Im analysing every poll and not only the last from every institute since 2004.

Link to post
Share on other sites
4 minutes ago, Entrecampos said:

Poll of Polls with information before 2008 election day. I think that will improve when I add 2004 polls for other states rather than Alabama, Alaska, Arizona and Arkansas + USA national polls
image.thumb.png.93fdc83d4aec3ef2b5fedcbab7255728.png
 

Is this to help with predicting 2020? If so, I doubt 2004 will help with that. If anything, distant elections are going to skew the likely outcome of 2020. Anything before 2012, probably shouldn't be included. I'd use 2012, then 2016 (weighed heavier than 2012), and then use average polling for 2020 (weighed heavier than 2016). Try this first without including 3rd parties and just look at margins, then make a second with 3rd parties. Average those two results together, and you'll probably get something closer to what the outcome will be. 

Link to post
Share on other sites
1 hour ago, vcczar said:

Is this to help with predicting 2020? If so, I doubt 2004 will help with that. If anything, distant elections are going to skew the likely outcome of 2020. Anything before 2012, probably shouldn't be included. I'd use 2012, then 2016 (weighed heavier than 2012), and then use average polling for 2020 (weighed heavier than 2016). Try this first without including 3rd parties and just look at margins, then make a second with 3rd parties. Average those two results together, and you'll probably get something closer to what the outcome will be. 

Yes, I will do weighed by time. Using that data for 2008 works very well like you can see here. 2004 would be a base for 2008, 2008 would be a base for 2012, 2012 would be a base for 2016, and 2016 would be a base for 2020, making a comparison with sports, tennis/nba/soccer ELOS work better than analysing one single season (ELO is better than ATP Tennis rankings for example). I think here will be the same, 2004 results alone are bad for predicting 2020 elections, but are a good base for 2008 polls that will be good for predict 2012 better than what would be if I didnt 2004 polls.
image.png.ca2333e7471a72596e08d3c390594af7.png

I disagree with that part, in the final, 2004 polls will have a very small importance, and only for 2020 active pollsters that exist since 2004, but I will skip them a bit for now . I Will do all states from 2016 before 2008, 2012 and 2004.

This will happen naturally with ELO model.

A very good idea! I will apply it.







 

  • Thanks 1
Link to post
Share on other sites

I will post here the results for now (model is far from complete) and will do the debut of States model, state by state.

Biden ups to 49.1% (+0.9%)
Trump downs to 42.8% (-1%)

Alabama

Trump 61.2%
Biden 36.2%

Alaska

Trump 52.9%
Biden  43.5%

Arizona

Trump 47.8%
Biden 49.6%

  • Upvote 1
Link to post
Share on other sites
29 minutes ago, Entrecampos said:

I will post here the results for now (model is far from complete) and will do the debut of States model, state by state.

Biden ups to 49.1% (+0.9%)
Trump downs to 42.8% (-1%)

Alabama

Trump 61.2%
Biden 36.2%

Alaska

Trump 52.9%
Biden  43.5%

Both of these state predictions are within the real of possibility for sure. 

Link to post
Share on other sites
40 minutes ago, vcczar said:

Both of these state predictions are within the real of possibility for sure. 

Arizona

Trump 47.8%
Biden 49.6%

Pennsylvania

Trump 45.9%
Biden 50.2%

I will do more tomorrow, but before I need to add more past polls to the database.

  • Like 1
Link to post
Share on other sites
12 minutes ago, Entrecampos said:

I need to decide if I will add midterm 2018 polls in the end or not... What are your opinions? Only presidential polls or other polls too?

That's a good question. Midterms 2010, in a way, forshadowed that Obama would have a harder time in 2012 than he did in 2008. 

I would have one model with it and one without it, and then average the two or something. It will probably be something like that. 

Link to post
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Loading...
×
×
  • Create New...