Slater's thread

[petm]

it doesn't do them 50 days in advance. It does them on the same day, different amounts in advance. e.g. on Aug 1st 31 days ahead for Sept 1st, 32 days ahead for Sept 2nd etc.

It does??

I don't think so. Could be wrong, but I don't think he implemented anything other than the 50-day projection. The SIPNs I checked for June and July are labelled "Slater - Persistence", which I assume is just anomaly persistence (the other line on the graph) not the model we are discussing. Only August SIPNs have the model prediction, at least the year I checked.

https://www.arcus.org/sipn/sea-ice-outlook/2014/june
https://www.arcus.org/sipn/sea-ice-outlook/2014/july
https://www.arcus.org/sipn/sea-ice-outlook/2014/august

[petm]

No, not if it is an artifact of the method used, in which case there is by definition no forecast for a minum.

Difference in interpretation...

[binntho]

No, not if it is an artifact of the method used, in which case there is by definition no forecast for a minum.

Difference in interpretation...

No, this is not interpretation. An artifact is an unintended result of a human action. If it is unintended it cannot be used to predict anything in a scientific context.

[petm]

It's a model output, not an artifact.

[binntho]

It's a model output, not an artifact.

Outputs can be intended or not intended. The model is built to deliver some specifics, and these are intentional. Looking at graphs of these specifics one might notice things that one didn't think the model was made to output. These are unintentional effects, for which I've used the word artifacts.

Unintentional effects may have value, but if you want to claim that they do, you have to make a good case for that claim. I've seen no such case, the "path" has obviously no predictive value and since the model is not predicting a minimum now (nor will we ever know when such a prediction might occur) then that goes the same way. The model does not, nor is it intended to, predict the path or the minimum.

[petm]

Personally I think it obvious that this is a matter of interpretation. But anyways this is not a very important point.

The main point is, and the one I was originally trying to make in the melting thread, that the Slater model looks like it is not predicting a record minimum extent.

[binntho]

Personally I think it obvious that this is a matter of interpretation. But anyways this is not a very important point.

The main point is, and the one I was originally trying to make in the melting thread, that the Slater model looks like it is not predicting a record minimum extent.

Agreed!

[El Cid]

Personally I think it obvious that this is a matter of interpretation. But anyways this is not a very important point.

The main point is, and the one I was originally trying to make in the melting thread, that the Slater model looks like it is not predicting a record minimum extent.

Furthermore, knowing its past record we can say that it forecasts a minimum somewhere below 4 M sq km and since its overestimate is almost always within (eyeballing) 0,5 M sq km it gives 2019 a solid second place finish and a very low chance of a first place. 

[oren]

And how can we tell when it is forecasting the minimum?

is by definition the model's forecasted minimum

No, not if it is an artifact of the method used, in which case there is by definition no forecast for a minum.

A minimum of the model's forecasts is NOT the forecasted minimum. It only sounds the same.

[El Cid]

And how can we tell when it is forecasting the minimum?

is by definition the model's forecasted minimum

No, not if it is an artifact of the method used, in which case there is by definition no forecast for a minum.

A minimum of the model's forecasts is NOT the forecasted minimum. It only sounds the same.

Then what is?

[GlassHalfEmpty]

It's sometimes worth considering simplifying to understand what is going on with something more complex (like the Slater model). So, here is my incredibly stupid model for predicting future ice extent.

Let's suppose that we have actual readings of x_i on day i. The prediction for day 50, in my model, is p_i=x_i+50(x_i - x_i-1).

Every day, this will make a prediction for 50 days in the future. Every day, we can plot that point, and we end up with a graph of p_i. It will form a curve (of sorts). At no point do any minima that this plot shows mean anything. Each underlying run is a linear model: apart from when the gradient is 0, the model can never predict a minima. The graph of p_i is completely meaningless in terms of predictions of minima.

Yes, Slater's model is much more complex: having said that, these basic principles of interpreting independent modelling runs still apply.

[binntho]

It's sometimes worth considering simplifying to understand what is going on with something more complex (like the Slater model). So, here is my incredibly stupid model for predicting future ice extent.

Let's suppose that we have actual readings of x_i on day i. The prediction for day 50, in my model, is p_i=x_i+50(x_i - x_i-1).

Every day, this will make a prediction for 50 days in the future. Every day, we can plot that point, and we end up with a graph of p_i. It will form a curve (of sorts). At no point do any minima that this plot shows mean anything. Each underlying run is a linear model: apart from when the gradient is 0, the model can never predict a minima. The graph of p_i is completely meaningless in terms of predictions of minima.

Yes, Slater's model is much more complex: having said that, these basic principles of interpreting independent modelling runs still apply.

Well put.

[oren]

Welcome, GlassHalfEmpty!

[El Cid]

Using your terms the Slater model in a perfectly certain world would be

x_i=y_i+z_i
p_i= y_i

where y is "thick" ice, which will never melt out in 50 days and z is "thin" ice which will always melt out, and x is total ice.

p in this case is a perfect prediction which gives you the exact future value for every given day and also gives you the path and the minimum.

In the real world the situation is the same, but we have no certainity just probabilities.

[Richard Rathbone]

it doesn't do them 50 days in advance. It does them on the same day, different amounts in advance. e.g. on Aug 1st 31 days ahead for Sept 1st, 32 days ahead for Sept 2nd etc.

It does??

I don't think so. Could be wrong, but I don't think he implemented anything other than the 50-day projection. The SIPNs I checked for June and July are labelled "Slater - Persistence", which I assume is just anomaly persistence (the other line on the graph) not the model we are discussing. Only August SIPNs have the model prediction, at least the year I checked.

https://www.arcus.org/sipn/sea-ice-outlook/2014/june
https://www.arcus.org/sipn/sea-ice-outlook/2014/july
https://www.arcus.org/sipn/sea-ice-outlook/2014/august

From the executive summary of the latest submission. https://www.arcus.org/files/sio/29668/nsidc_barrettslaterpan-arctic.pdf


"Forecasts issued on July 1 for September have lead times spanning 62 to 91 days."

Looking through the history it seems this is an innovation of Barrett and Slater did indeed use multiple starting days at the same lead time while Barrett uses the Slater model with one starting day at multiple lead times.

[DrTskoul]

If that is not a decent model for projecting extent minimum , I don't  know what is

[binntho]

Well it shows a strong correlation between the minimum of the Slater model and the observed minimums. How could it be otherwise? The Slater model is based purely on statistics and averages, and does not make predictions above and beyond that.

So we could say that if everything stays within the statistics, Slater makes a decent forecast. But as 2012 shows very clearly, it can't catch anything outside the statistics.

Still, given all my slurs about the supposed predictive abilities of the Slater model, I do concur (and have already conceded) that the Slater model does give a very good indication of the magnitude of the minimum when we get this close.

And the Slater model is very useful and interesting and informative - but metaphysical attempts to scry from the graph more than it can deliver goes against my sensibilities.

[DrTskoul]

Model performance last three years. Source:
Slater Probabilistic Ice Extent - Cooperative Institute for Research in Environmental Sciences

Edit: added link.

[petm]

From the executive summary of the latest submission.
Looking through the history it seems this is an innovation of Barrett and Slater did indeed use multiple starting days at the same lead time while Barrett uses the Slater model with one starting day at multiple lead times.

Interesting. Too bad they don't publish this on the website.

[binntho]

Model performance last three years. Source:
Slater Probabilistic Ice Extent - Cooperative Institute for Research in Environmental Sciences

Very interesting (and a link would have made it even better) - so Slater's model (which I presume is much simpler than the other models in this comparison) has a very good track record when it comes to predicting the magnitude of the minimum.

So it seems I might have to consider a total retreat from previously held positions ...

[S.Pansa]

That coulb be the Link i guess.
No, only schows the main page. But you can find the comparisons on the left, under "Model Performance"

[binntho]

But I'd want to maintain that Gerantocrat's daily "Average remaining melt would ..." predictions are even better.

But how to compare? Gerantocrat updates his prediction for the minimum daily, Slater doesn't. So Slater's current prediction of very close to 4 MKm2 was made on the 14th of July. That day, Gerantocrat predicted a minimum of 4.05 Mkm2.

But since then Gerantocrat has continued to make preditions, today he is predicting 3.80 MKm2 based on pure statistics and nothing else, so since mid July the two "models" have diverged considerably (sorry G. for calling you a model).

So all in all I'd go with the G model.

[El Cid]

Slater's model had a minimum of 3,85 M sq km

[binntho]

Slater's model had a minimum of 3,85 M sq km

Where did you see that? I'm seeing practically exactly 4Mkm2

[gerontocrat]

But I'd want to maintain that Gerantocrat's daily "Average remaining melt would ..." predictions are even better.

But how to compare? Gerantocrat updates his prediction for the minimum daily, Slater doesn't. So Slater's current prediction of very close to 4 MKm2 was made on the 14th of July. That day, Gerantocrat predicted a minimum of 4.05 Mkm2.

But since then Gerantocrat has continued to make preditions, today he is predicting 3.80 MKm2 based on pure statistics and nothing else, so since mid July the two "models" have diverged considerably (sorry G. for calling you a model).

So all in all I'd go with the G model.

Change "the model predictions ...." to "the spreadsheet projections" and I am content.

Saying "the model predicts" carries with it a flavour of approval, agreement. I used to add a sentence something like:-

The accuracy of the data as presented has been checked.
The content does not necessarily reflect the opinion of the author.

I am still sticking with a 3.5 to 4 million km2 guess with a bias to 3.75 to 4 million km2. Which could be totally upended either way by the PIOMAS data (due in a week's time?).
________________________________________________

Gerantocrat

Gerontocrat

/dʒɛˈrɒntəkrat/
noun
a member or leader of a state, society, or group governed by old people.
"the village's aged leader is portrayed as a dictatorial gerontocrat"
A Grumpy old man with power...   my personal Nirvana

[DrTskoul]

Γεροντοκράτορας....

[binntho]

I stand corrected!

[gerontocrat]

I stand corrected!

I've just been promoted by DrTskoul.

Translation from Google....

Γεροντοκράτορας                           Elder Emperor

Where's me crown?

[binntho]

Change "the model predictions ...." to "the spreadsheet projections" and I am content.

Well, are the models anything more than glorified spreadsheet projections? There's no magic involved, as far I'm aware.

The difference between the Slater model and the G projections are that the G is predicting the minimum every day, while we can only see with hindsight which day the S predicted the minimum.

[aslan]

I stand corrected!

I've just been promoted by DrTskoul.

Translation from Google....

Γεροντοκράτορας                           Elder Emperor

Where's me crown?

Off topic, but it is as old as the human societies זִקְנֵי הָעָמ or זִקְנֵ֥י יִשְׂרָאֵ֖ל

[DrTskoul]

Change "the model predictions ...." to "the spreadsheet projections" and I am content.

Well, are the models anything more than glorified spreadsheet projections? There's no magic involved, as far I'm aware.

Ok, there... D-K in full blown display...

[binntho]

Change "the model predictions ...." to "the spreadsheet projections" and I am content.

Well, are the models anything more than glorified spreadsheet projections? There's no magic involved, as far I'm aware.

Ok, there... D-K in full blown display...

Well speaking as a computer programmer (who should probably be doing his work right now) then, yes, every computer model or simulation is, when it comes down to it, just manipulating numbers according to well-defined rules. A glorified spreadsheet, in other words.

[gerontocrat]

Change "the model predictions ...." to "the spreadsheet projections" and I am content.

Well, are the models anything more than glorified spreadsheet projections? There's no magic involved, as far I'm aware.

Ok, there... D-K in full blown display...

Well speaking as a computer programmer (who should probably be doing his work right now) then, yes, every computer model or simulation is, when it comes down to it, just manipulating numbers according to well-defined rules. A glorified spreadsheet, in other words.

every computer model or simulation is a glorified spreadsheet
I completely disagree.

For the projections on the minimum I put out simply take the data as it is and plonk it into a table..
BUT:-
When I did  spreadsheets for financial projects, I often put in a sheet called "parameters".
e.g....
How long will the project take to build?
How much will it cost?
How much will I have to borrow and on what terms?
Will I be able to sell the product it produces (e.g. potable water) at a price that will recover the costs?
Will there be savings elsewhere (manpower, replacing old inefficient facilities)?

The same can apply to environmental energy projects, changing the £ or $ sign to KwH (i.e. from financial equations to energy equations).

Imagine the conversation between me, who has seen too much of this stuff, and the proposer (private, World Bank, a Government officer) desperate to make their pet project happen.

The result ranging from a complete bust to a magnificent proposal.

BIAS, I think it is called. That is the difference between a spreadsheet projection and a model prediction.

But of course, bias never ever creeps into parameter driven climate models, does it.
But of course, bias never ever creeps into parameter driven geo-engineering proposals, does it.

If I take my spreadsheet and start putting in parameters on:-
- general increase in Arctic temperatures,
- effect of increased ice mobility,
- the modest El Nino,
- etc etc
then they reflect opinion as well as fact, and so my spreadsheet projections become model predictions, and I must make that clear when publishing them.

Maybe off-topic but to me it matters a lot. There is a whole thread in the ASIF mainly on whether IPCC projections contain bias (conscious or subconscious or unconscious).

[binntho]

I've worked with quite advanced spreadsheets myself, though probably not as advanced as yours, Gerontocrat, and I've worked with some pretty advanced programming as well and to me there is no fundamental difference.

Most models and simulations use some form of matrix manipulation, which is even looks like spreedsheets!

But never mind, I didn't really mean anything very profound by that "glorified spreadsheet" comment - other than saying that all they ever do is manipulate numbers according to predefined rules.

Having said that (like the man said), would you still call Slater's model a "model"? As far as I can see his model is not doing anything fundamentally different from what yours is doing. You work on aggregates, the Slater model is more finegrained and involves some statistics, but at the end of the day it's just making predictions (or projections if you like) based purely on extrapolation of historical data combined with whatever state the ice is in today.

[binntho]

As to whether a model is without bias - isn't that the same as asking a journalist if he is objective? Both are humanly impossible.

[El Cid]

Slater's model had a minimum of 3,85 M sq km

Where did you see that? I'm seeing practically exactly 4Mkm2

I used to post the Slater results on the main meting thread. I posted this on July 14:

Slater update

The model has taken a dive, forecasting 3,87 M sq km for Sep 2, 2019

as we know only in 2012 did NSIDC extent dive under the 4 M sq km mark

For comparison, extent (in M sq km) on Sep 2 of

2012: 3,56
2016: 4,29
2007: 4,46

So based on this 2019 is almost guaranteed a second place, although I would not be surprised by a record as the model overestimated extent during all strong (preconditioned? melt mometum?) years (2007,12,16) by cca 0,5 M sq km. 

[DrTskoul]

Mathematical models can take many forms, including dynamical systems, statistical models, differential equations, or game theoretic models. These and other types of models can overlap, with a given model involving a variety of abstract structures. In general, mathematical models may include logical models. In many cases, the quality of a scientific field depends on how well the mathematical models developed on the theoretical side agree with results of repeatable experiments. Lack of agreement between theoretical mathematical models and experimental measurements often leads to important advances as better theories are developed.

In the physical sciences, a traditional mathematical model contains most of the following elements:

Governing equations
Supplementary sub-models
Defining equations
Constitutive equations
Assumptions and constraints
Initial and boundary conditions
Classical constraints and kinematic equations

A model can be solved on the back of a napkin, or with an excel spreadsheet or a supercomputer. A model is a mathematical representation of a physical system. Can be as simple as a single equation or as complicated as the quantum mechanical simulations in the Standard model.

[binntho]

Slater's model had a minimum of 3,85 M sq km

Where did you see that? I'm seeing practically exactly 4Mkm2

I used to post the Slater results on the main meting thread. I posted this on July 14:

Slater update

The model has taken a dive, forecasting 3,87 M sq km for Sep 2, 2019

as we know only in 2012 did NSIDC extent dive under the 4 M sq km mark

For comparison, extent (in M sq km) on Sep 2 of

2012: 3,56
2016: 4,29
2007: 4,46

So based on this 2019 is almost guaranteed a second place, although I would not be surprised by a record as the model overestimated extent during all strong (preconditioned? melt mometum?) years (2007,12,16) by cca 0,5 M sq km.

I remember when you used to post these - but where do you get that figure from? I'll admit that the quality of the graphic is such that 3.87 is well within reason, but what is the source of the specific figures?

EDIT: Is it the figure at the bottom of the map image?

[GlassHalfEmpty]

Welcome, GlassHalfEmpty!

It's good to be here I've been lurking (without an account) for quite a few years, and thought that I may as break cover when the subject hit something that I have some experience with (probabilistic modelling/computer simulation).

Anyway, since I'm not a big fan of off-topic posting...

What can it mean when a 50-day predictions show an increase? The answer is: it depends.

To keep the analysis below a little bit simpler, I'm going to assume that the general shape of each day's model is smooth and bowl-shaped (basically, a bit like a parabola), and that we are looking at predictions at, or before, each day's model predicts there is a minima. Don't forget that each day has a completely new model generated, and each of these models will have its own shape (and minima, both level and date).

There are (to me) three obvious reasons why we might see an upturn in the prediction graph:

• Option 1: this is a prediction of an actual increase.
  Today's model, and yesterday's model, are basically the predicting the same future, and that future has an increase.
• Option 2: the prediction is that the future minima is at the same time, but is higher.
  If the model's minima is expected to be on the same day, but has had its value increased, today's 'bowl' will be shallower than yesterday's, and so dates on the way to it will (generally) have higher values today than they did yesterday - this includes day 50, and it is quite possible for the curve to have shallowed sufficiently that today's day 50 is not only higher than yesterday's day 51, but also it's day 50.
  
• Option 3: the prediction is that the future minima is later.
  Same as Option 2, except this time the 'bowl' is wider - the path down to the minima (as for Option 2) has a slower rate of descent, so the same reasoning will apply. Because it is later, the new model's minima could also actually be a lower value.
  

(Of course, the real model is not likely to be quite so well-behaved in terms of being 'bowl'-shaped, so it will almost certainly have predictions of local minima, but that output isn't published, so it is hard to tell.)

The more years of historical data that the underlying statistical model has available, the more likely it is that 'Option 1' will (approximately) hold... but only as long as the current year follows the same 'rules' as previous years. But there is no reason why, as new data comes in that is away from the average, that each day's model doesn't revise expected level/date of the actual model minima. And that could have surprising consequences (Options 2 and 3.)

Two things I wish I could see from the model: the uncertainty for each prediction - it feels like the model must have that value available - and the gradient at each day 50.

(Oh, yeah, Option 4: all the previous models had estimated a later minima, and the later model runs then push the minima to an earlier date (before day 50), so that now the predictions are for the other side of the 'bowl'. Seems fairly unlikely in this case.)

[blumenkraft]

Welcome to the forum GlassHalfEmpty.

Let me just say, imho you are using the word prediction a little bit too loosely. Could we agree on the word 'forecasted' in this case?

I would say a prediction can not be done with so many unknowns.

[El Cid]

Slater's model had a minimum of 3,85 M sq km

Where did you see that? I'm seeing practically exactly 4Mkm2

I remember when you used to post these - but where do you get that figure from? I'll admit that the quality of the graphic is such that 3.87 is well within reason, but what is the source of the specific figures?

EDIT: Is it the figure at the bottom of the map image?

I looked at the site (slater's) every day and posted the numbers I saw on ASIF

[GlassHalfEmpty]

Let me just say, imho you are using the word prediction a little bit too loosely. Could we agree on the word 'forecasted' in this case?

I would say a prediction can not be done with so many unknowns.

I'm easy. I would quibble that it is less to do with how many unknowns there are on more to do with the fact that we are dealing with time series data, and so "forecast" is probably more common.

This is probably why I generally prefer writing things out mathematically - it's the evaluation of a model at time t, and make of the value what you will

[petm]

The model has taken a dive, forecasting 3,87 M sq km for Sep 2, 2019

From the following day, which was a touch higher.

(I have a script that downloads the map every day, but sometimes it fails. I don't have Sept. 2.)

[petm]

The Slater model seems to have stabilized at ~4.3M km2. The Sept. average prediction forecast will probably be near this value, , roughly tied for 2nd place (in terms of forecast) with 2011 (see https://forum.arctic-sea-ice.net/index.php/topic,1682.msg217704.html#msg217704). As discussed in detail earlier in this thread, it is the September average value that the model has good skill at predicting, not the 1-day minimum. The most it has overestimated so far is ~0.5M km2 (e.g., in 2012). If that holds this year, we shouldn't expect the Sept average to be much below 4M at the lowest. We shall see...

[binntho]

Well argued, GlassHalfEmpty, and I agree with petm that the Slater model does have a very good track record for predicting September average (thanks for reminding us).

[binntho]

Googled "prediction vs. forecast" and got this:

Forecast is scientific and free from intuition and personal bias, whereas prediction is subjective and fatalistic in nature. Forecasting is an extrapolation of past into the future while prediction is judgmental and takes into account changes taking place in the future.

[Richard Rathbone]

I thought I should check the recent accuracy, seeing as the method for the SIPN entries changed slightly under Barrett and the comparisons already posted end several years earlier.

SIPN forecasts and results in the Slater/Barrett era. I think these were two easy years to get right (because close to trend), but the results were very close to the model.

2018: Sept average 4.71, predictions Aug 4.75, July 4.97
2017: Sept average 4.80, predictions Aug 4.77, July 4.82

Unfortunately I don't get to see the SIPN entry until after forum poll deadlines, so I'm guessing based on the 50 day forecasts around now, but its pretty hard to bet against what the Slater model comes up with on Aug 1st.

[blumenkraft]

Googled "prediction vs. forecast" and got this:

Forecast is scientific and free from intuition and personal bias, whereas prediction is subjective and fatalistic in nature. Forecasting is an extrapolation of past into the future while prediction is judgmental and takes into account changes taking place in the future.

I don't understand that.

I'm not a native English speaker, but in my language, a prediction is something like this: When i tip the pencil over the edge of the table, it will fall to the floor. (assuming there are no other things between the floor and the tabletop).

This is not fatalistic nor is it subjective, or is it?

[binntho]

Googled "prediction vs. forecast" and got this:

Forecast is scientific and free from intuition and personal bias, whereas prediction is subjective and fatalistic in nature. Forecasting is an extrapolation of past into the future while prediction is judgmental and takes into account changes taking place in the future.

I don't understand that.

I'm not a native English speaker, but in my language, a prediction is something like this: When i tip the pencil over the edge of the table, it will fall to the floor. (assuming there are no other things between the floor and the tabletop).

This is not fatalistic nor is it subjective, or is it?

Well neither am I a native speaker, and when you told GHE off about using prediction a bit freely, I became intrigued.

Personally, I've always understood forecasts as belonging specifically to meteorology, and when used generally, then strictly in the meaning of extrapolating current trends.

On the other hand, predictions cover both forecasts, prophecies and prognostications as far as I understand it, and is generally used extremely loosely, such as when we say that Slater's model "predicts" things.

As with all models, do they ever predict anything? They make extensive and at times expansive forecasts but, without truly prophetic abilities, it's impossible to know what the future holds, and it is up to us as individuals to make predictions, such as "I predict that the September average monthly minimum is going to be around 4Mkm2".

[blumenkraft]

As with all models, do they ever predict anything?

No. They estimate, they calculate probabilities, they take knowns into account and they are unaware of black swans; and that's it.

They make extensive and at times expansive forecasts but, without truly prophetic abilities, it's impossible to know what the future holds,

Well, that's the thing. I can predict this pencil will fall, 100% of the time. It will always work, even in the future.

Or when Einstein thought about relativity and was able to predict gravitational lenses. And by doing so, he proved his theory to be correct. He made the claim, that you can see stars behind the sun during a solar eclipse.

You can make true claims about the future. And those are the ones that would fall in the category of predictions.

Sorry for being pedantic here, but is my understanding of the word wrong?