Land snow cover effect on sea ice

[bbr2314]

OK, to be specific, let's say that we take the average of the 500 mb geopotential height anomaly over the province of Quebec (that's entirely within your blue region, and covers about half of it, and there's no dispute about the boundaries) for the month of June.

You predict -100m. CFSv2 predicts about +20m.

Sound like a reasonable and measurable prediction to you?

Time will tell who gets closer.

Reasonable!

[Dharma Rupa]

OK, to be specific, let's say that we take the average of the 500 mb geopotential height anomaly over the province of Quebec (that's entirely within your blue region, and covers about half of it, and there's no dispute about the boundaries) for the month of June.

You predict -100m. CFSv2 predicts about +20m.

Sound like a reasonable and measurable prediction to you?

Time will tell who gets closer.

Reasonable!

Can we set up a thread specifically for this prediction, and stop discussing this matter everywhere else?

[jdallen]

So you predict record cold June in the Kivalliq, Northern Ontario, all of Quebec, the Maritimes and Maine (among other regions).

How do we measure that exactly? Average temp, nightly lows, daily highs?

<snippage>
PS: Welcome to May! If the 12z EURO is correct, even WITH its recent warm bias, there is no end to winter in sight for some regions. Just temporary breaks. It will be very interesting to see how much SWE melts off in May, as if it may indicate an upper bound of possibility (or not). I do not think it will be more than 750KM^3 (for North America). I suspect Eurasia will bear the brunt of the May melt, possibly even making things more lopsided than today (however temporarily). 

bbr2314, I'm having a serious credulity issue with a lot of what you are suggesting.

[bbr2314]

Does anyone know if there is a data source comparable to PIOMAS for accumulated extant hemispheric SWE? If so, I would be very curious to see a running % comparison for the total amount of hemispheric sea ice AND SWE contained within the SWE. My rudimentary math says we are up from 7.5% to 13.5% from the early 00s to 2018. I wonder what year the shift to increasing % of SWE began? I suspect it was 1997.

[gerontocrat]

I have been looking at Hudson Bay a bit.
For this topic = "Re: Land snow cover effect on sea ice" it seems to be a nice little test bed, because:-

During winter and until May at least Hudson Bay is more or isolated from the rest of the Arctic Ocean.
In addition the snow thickness data shows it surrounded by loads of snow this winter.

So what happened - yes, ice area was right up there until about mid-April. But now the melt is happening even though snow has not melted out in that part of the world. The next week or three should be interesting.

[Archimid]

I think the snow cover over Alaska is helping the the ice in the Chukchi. I think that because when I look at nullschool wind+temps at surface the temperature is at or near freezing levels.

https://earth.nullschool.net/#current/wind/surface/level/overlay=temp/orthographic=-168.33,60.51,674/loc=-155.357,65.389

However when I look at 1000hPa it's mostly above freezing.

https://earth.nullschool.net/#current/wind/isobaric/1000hPa/overlay=temp/orthographic=-168.33,60.51,674/loc=-155.357,65.389


Why is it that much colder in the surface? I think that's because of snow. Snow serves the purpose of cooling down the air that's being kinetically pushed into the Chukchi. If the same air flowed through open land then the the air arriving to the Chukchi would be much warmer.

This protection is quickly running out.

[Coffee Drinker]

Isn't the Hudson Bay still 100% frozen over? How can the extend then vary at this time of the year? Shouldn't it be a straight line (full cover) until the first open water appears?

[Neven]

Isn't the Hudson Bay still 100% frozen over? How can the extend then vary at this time of the year? Shouldn't it be a straight line (full cover) until the first open water appears?

Nope, not 100%:

[jdallen]

Isn't the Hudson Bay still 100% frozen over? How can the extend then vary at this time of the year? Shouldn't it be a straight line (full cover) until the first open water appears?

Active melt has begun and there are large areas of both open water and extensive fracturing.

Image from Wordview on the 21st, which of recent days had the best visibility.

[ghoti]

Look at the James Bay portion of Hudsons on the 20th-25th. It is not just pulling away from the coast but melting out pretty fast.

[bbr2314]

I wonder what the impact of this will be. That is a lot of runoff at once and it could continue for a few days. Hemispherically, also major.

[Iceismylife]

I wonder what the impact of this will be. That is a lot of runoff at once and it could continue for a few days. Hemispherically, also major.

Ice jams. Slow down the overturning?  Flooding?

[Coffee Drinker]

Isn't the Hudson Bay still 100% frozen over? How can the extend then vary at this time of the year? Shouldn't it be a straight line (full cover) until the first open water appears?

Active melt has begun and there are large areas of both open water and extensive fracturing.

Image from Wordview on the 21st, which of recent days had the best visibility.

Thank you. That was quite surprising for me considering the temperatures and the typical resilience of sea ice. There have been really only a handful of warm (>+10C) days so far and nights are still quite frosty. 

[RikW]

Yeah, that should be a lot of melt-water. And a lot more than normal, especially if it continues to melt this fast.

[litesong]

Isn't the Hudson Bay still 100% frozen over? How can the extend then vary at this time of the year? Shouldn't it be a straight line (full cover) until the first open water appears?

Active melt has begun and there are large areas of both open water and extensive fracturing.

Image from Wordview on the 21st, which of recent days had the best visibility.

Thank you. That was quite surprising for me considering the temperatures and the typical resilience of sea ice. There have been really only a handful of warm (>+10C) days so far and nights are still quite frosty.

Yes, Canada (especially northern Canada & its archipelago) have been anomalously very cold this past winter, often rivaling the North Pole in temperatures. A vast cold system extending from Canada, across the North Pole & to mid-Siberia & northern China was persistent. It is the reason that what I called the Present High Arctic Berserker(2) (continuous winter time, High Arctic over-temperatures), PHAB (2), or FAB (2) truncated at "only" 215 days, not matching or breaking the record FAB (1) of 230+ days in latter 2016 to first part of 2017 time period.

[numerobis]

Yes, Canada (especially northern Canada & its archipelago) have been anomalously very cold this past winter, often rivaling the North Pole in temperatures. A vast cold system extending from Canada, across the North Pole & to mid-Siberia & northern China was persistent. It is the reason that what I called the Present High Arctic Berserker(2) (continuous winter time, High Arctic over-temperatures), PHAB (2), or FAB (2) truncated at "only" 215 days, not matching or breaking the record FAB (1) of 230+ days in latter 2016 to first part of 2017 time period.

December was cold over Western Hudson Bay, but otherwise was scorching across the Canadian Arctic.

January was warm over most the Canadian Arctic, except for Eastern Hudson Bay and Baffin.

February was quite cold over the lower Canadian Arctic, but scorching over the High Arctic. It was frequently much warmer in Alert than in Iqaluit that month.

March was scorching over Nunavik and southern Baffin, warm over most of the Arctic.

So I'm not sure your "anomalously very cold" idea holds water.

[litesong]

December was cold over Western Hudson Bay, but otherwise was scorching across the Canadian Arctic. January was warm over most the Canadian Arctic, except for Eastern Hudson Bay and Baffin. February was quite cold over the lower Canadian Arctic, but scorching over the High Arctic. It was frequently much warmer in Alert than in Iqaluit that month.
 March was scorching over Nunavik and southern Baffin, warm over most of the Arctic.
So I'm not sure your "anomalously very cold" idea holds water.

  Yes, a winter time system of similar temperature ranging from Canada, to the North Pole & south to central Siberia & northern China reigned. If talking about Canada, Siberia or northern China, those "anomalous temperatures " were "cold". Those same temperatures, if talking about the High Arctic, were "hot". Through the course of the winter, that pattern of similar temperature did fluctuate around a bit, letting warm southern heat in occasionally, even to the North Pole. 

[ghoti]

The DMI 80N temperatures were above the 30 year mean the whole time. You have a different definition of anomalously cold it seems.

[Neven]

I also don't understand what the 'anomalously very cold' is based on. Here's a temperature anomaly map (925 hPa) for the Arctic from Nov 1 to Mar 31:

[litesong]

The DMI 80N temperatures were above the 30 year mean the whole time. You have a different definition of anomalously cold it seems.

Quoting High Arctic temperatures, in which I agree that PHAB (2) or FAB (2) was above normal for 215 straight days, is NOT the way to disprove that Canadian temperatures were low this winter. Even now, a much smaller central portion of north Canada is much lower than normal. Through much of the winter, large areas of Canada were sub-normal cold.

[Neven]

Thanks Neven, for proving my contention that Canada averaged cold this winter, while the High Arctic was warm.

Oh, I'm sorry. It was this part that stood out for me: "Canada (especially northern Canada & its archipelago) have been anomalously very cold this past winter. A vast cold system extending from Canada, across the North Pole & to mid-Siberia & northern China was persistent."

I didn't know that meant "Canada averaged cold this winter". 

[jdallen]

I also don't understand what the 'anomalously very cold' is based on. Here's a temperature anomaly map (925 hPa) for the Arctic from Nov 1 to Mar 31:

Thanks Neven, for proving my contention that Canada averaged cold this winter, while the High Arctic was warm.

Careful. Let's keep it civil, for your good.

ERSL shows D-180 average anomaly over the prairie provinces up into southern NW Territories being about 1 degree below normal, with a bit colder over James bay, with a max anomaly of -2.5.

It's not particularly impressive actually, and reflects the effect of multiple arctic breakouts this winter across the Canadian shield.  Adding in the parts which were above to extremely above normal, I'd actually say it evens out to show Canada overall was about average.

https://www.esrl.noaa.gov/psd/map/images/rnl/sfctmpmer_180b.rnl.gif

The 90 day average has the anomaly further west mostly over Saskatchewan and Alberta with a tail running NW over NW British Columbia and the southern Yukon.  Once again, this is reflective of arctic breakouts and doesn't show the country as a whole being below average.  Hudson's bay in particular isn't affected in this plot.

https://www.esrl.noaa.gov/psd/map/images/rnl/sfctmpmer_90b.rnl.html

The 30 day average does show substantial downward anomalies over most of the country, but again, that's reflective of this season's breakouts and is more noise than trend.

https://www.esrl.noaa.gov/psd/map/images/rnl/sfctmpmer_30b.rnl.html

[Dharma Rupa]

It's not particularly impressive actually, and reflects the effect of multiple arctic breakouts this winter across the Canadian shield.  Adding in the parts which were above to extremely above normal, I'd actually say it evens out to show Canada overall was about average.

Thank you.  This comports with my expectations that WACCy weather will only mean cold contenents in comparison to the ocean -- not in absolute terms.  They might get more winter snowfall because of the increased humidity, but I don't expect that to have any significant effect.

[numerobis]

December was cold over Western Hudson Bay, but otherwise was scorching across the Canadian Arctic. January was warm over most the Canadian Arctic, except for Eastern Hudson Bay and Baffin. February was quite cold over the lower Canadian Arctic, but scorching over the High Arctic. It was frequently much warmer in Alert than in Iqaluit that month.
 March was scorching over Nunavik and southern Baffin, warm over most of the Arctic.
So I'm not sure your "anomalously very cold" idea holds water.

  Yes, a winter time system of similar temperature ranging from Canada, to the North Pole & south to central Siberia & northern China reigned. If talking about Canada, Siberia or northern China, those "anomalous temperatures " were "cold". Those same temperatures, if talking about the High Arctic, were "hot". Through the course of the winter, that pattern of similar temperature did fluctuate around a bit, letting warm southern heat in occasionally, even to the North Pole.

Can you show data that backs up what you're saying? You make it sound like we agree, but if so, I'm very much not understanding you.


Besides, what does this have to do with the effect of land snow cover upon sea ice?

[litesong]

Can you show data that backs up what you're saying?

I don't have to. jdallen already did in his linked months long global temperature graph:
https://www.esrl.noaa.gov/psd/map/images/rnl/sfctmpmer_180b.rnl.gif
Note the anomalous low temperatures over Canada.

[jdallen]

Can you show data that backs up what you're saying?

I don't have to. jdallen already did in his linked months long global temperature graph:
https://www.esrl.noaa.gov/psd/map/images/rnl/sfctmpmer_180b.rnl.gif
Note the anomalous low temperatures over Canada.

You need to recall I presented that with caveats, m'friend.

[numerobis]

That map contradicts your claim.

The Canadian Arctic runs hot everywhere, except for southwestern Hudson Bay.

[jdallen]

From elsewhere...

Well, a quick look at the literature shows that snow over no-shrub tundra has the highest albedo, but also melts out much faster come spring. With increased vegetation, from dwarf shrub to low shrub to conifer forests, albedo drops significantly.In April (with snow cover), boreal forests have albedo of around 0.4, tree-shrub-mosaic tundra about 0.5, and other types of tundra between 0.72 and 0.76. Fresh snow has albedo of 0.8-0.9.

On longer timescales, since boreal forests and shrub cover is rapidly moving northwards, one would expect the effects of snow cover on albedo to drop as well.

<snip>

Thumbnail calculation - to melt 1 meter of snow/water equivalent snow cover requires about 33 million joules or about 9.3 KWH equivalent of heat.  This time of year, we're getting right around 6.5KWH/day of heat at the Arctic circle, so even if we were to double the amount of snow cover, we're still only talking about a day and a half's additional insolation to melt it - not that big an effect.

Even if we do have double the normal snow cover in thickness, that doesn't change albedo, and as binntho points out snow cover under boreal forests and taiga isn't has high as it is on tundra.

The real question is not whether more snow cover changes things, its whether or not increased albedo changes the system dynamics.  If we can sketch out regionally what the changes in albedo were, localized year over year, then we can compute the relative increase or decrease in seasonal heat uptake.

Otherwise, our arguments over snow cover is going to be just so much hand waving and anecdote, as we aren't discriminating between snow depths at varying latitudes, or between tundra, taiga, boreal forest or other biome.  A single large scalar value for snow cover doesn't really help us be more skillful in understanding the influence on the melt season.

[numerobis]

This map shows when the melt is happening:
http://nsidc.org/soac/snow-cover.html#snowcover

It plots how many days were snow-covered relative to average. Where there's anomalies is by definition where it's melting out. The melt line doesn't normally reach the shores of the Arctic until June, and the high Arctic not until July.

[ghoti]

The impact of all this extra snow on the sea ice may or may not be noticed. However, the impact of the melting of this extra snow now is flooding. Major floods in several regions of Canada this spring.

[Rob Dekker]

Most of you will know that since 2013, I use I use the "whiteness" of the Arctic in June as a predictor for how much ice will melt out between June and September.

http://neven1.typepad.com/blog/2013/07/problematic-predictions-2.html

Specifically, I use three variables to make this prediction :
- Land snow cover in June
- Ice 'area' in June
- (Extent - Area) in June, which represents the amount of 'water' in the ice pack in June.

A combination of these variables, each one of which affects the 'albedo' of the Northern Hemisphere, represents how much solar energy gets absorbed by the Northern Hemisphere in summer, and this correlates remarkably well with September sea ice cover.

Details of this method is described in one of my entries into Arcus Sea Ice Prediction Network :

https://www.arcus.org/files/sio/25738/sio-2016-july_dekker.pdf

This year, land snow cover in June was quite high compared to recent years :



Also, ice 'area' is quite high in June (in between 2014 and 2015) and the ice is still fairly compact.

As a result, prediction for Sept 2018 September sea ice extent is quite high at 5.19 km², with a standard deviation of 340 k km².

My gut feeling this year tells me that this an upper bound, but it's fairly clear that given the past performance of this method, it is highly unlikely (less than 2.5% chance) that Sept 2018 will end up below 4.5 M km².

Here is what this hind-cast method did for the past 26 years :

[Ned W]

Rob, thanks for reminding us about that.  It's pretty impressive.  I'll be curious to see how it works out this year.

[Ned W]

Are the right-most two green triangles the predictions for 2017 and 2018?  The predictions seem to be very close to each other, even though on the bar graph of snow cover anomalies there's a fairly large difference between the two years.  Does that mean the effect of the difference in snow cover was balanced by some combination of the other two variables in the model?  (That probably is a dumb question.)

[Rob Dekker]

Are the right-most two green triangles the predictions for 2017 and 2018?  The predictions seem to be very close to each other, even though on the bar graph of snow cover anomalies there's a fairly large difference between the two years.  Does that mean the effect of the difference in snow cover was balanced by some combination of the other two variables in the model?  (That probably is a dumb question.)

Yes. That is correct.
The other two variables that compensated for the snow cover are sea ice 'area' and compaction (I use 'extent minus area').

In the table below, from here :
ftp://sidads.colorado.edu/DATASETS/NOAA/G02135/north/monthly/data/N_06_extent_v3.0.csv

you can see that last year 'area' was lower than this year, and also the ice is more compact than last year :

2014,  6,      Goddard,      N,  11.03,   8.76
2015,  6,      Goddard,      N,  10.88,   8.70
2016,  6,      NRTSI-G,      N,  10.35,   8.13
2017,  6,      NRTSI-G,      N,  10.72,   8.56
2018,  6,      NRTSI-G,      N,  10.71,   8.71

[Ned W]

Thanks.  Sorry, but your reward for answering one question is getting asked another ... Is that "340k km2" simply the standard error of the regression model (i.e., standard deviation of the residuals) or is it a prediction interval?  This year I've started providing "prediction intervals" (wider than the traditional confidence interval) with the predict-o-matic predictions, which is appropriate since they are in fact ... predictions.

[Rob Dekker]

Thanks.  Sorry, but your reward for answering one question is getting asked another ... Is that "340k km2" simply the standard error of the regression model (i.e., standard deviation of the residuals) or is it a prediction interval?  This year I've started providing "prediction intervals" (wider than the traditional confidence interval) with the predict-o-matic predictions, which is appropriate since they are in fact ... predictions.

That 340 k km² is the standard deviation of the residuals.
Which translates to the 95% certainty prediction interval (assuming the distribution is Gaussian) being +/- 680 k km².

That's why I stated that there is only a 2.5% chance of this Sept extent going below 4.5 M km²

[Ned W]

I am not a statistician.  But my understanding is that what you're reporting is the confidence interval, which represents the "population", rather than the prediction interval, which represents the uncertainty around a prediction of a single new value.  This explanation here might be useful:

https://onlinecourses.science.psu.edu/stat501/node/274/

particularly the contrasting formulas for confidence interval and prediction interval, halfway down the page.

Again, I'm not a statistician, and my familiarity with this is relatively shallow.

[Ned W]

Or this is probably the equivalent explanation for a multiple linear regression model:

https://onlinecourses.science.psu.edu/stat501/node/315/

Also, and more troublingly, I don't know about your model, but for mine I am increasingly dubious that the distribution of errors is gaussian.  I think the distribution may be wider on the downside.  So although I have been reporting the prediction interval, I am becoming skeptical of its validity.  I should probably look into that.  Lately I have been more interested in examining the uncertainty than in the estimate itself, so I probably ought to be sure I'm getting it right.

[Rob Dekker]

Lately I have been more interested in examining the uncertainty than in the estimate itself, so I probably ought to be sure I'm getting it right.

I'm not a statistician either, so please let me know what you find out.

[Steven]

As a result, prediction for Sept 2018 September sea ice extent is quite high at 5.19 km², with a standard deviation of 340 k km².

I tried to reproduce your calculation, using the Rutgers and NSIDC data for June.  I reproduced your prediction of 5.19 M km2 for the September 2018 NSIDC extent.  But for the standard error of the regression, I calculated 0.38 M km2 rather than 0.34 M km2.

I guess you calculated the standard deviation of the residuals like this:


where e₁, e₂, ..., e_n are the residuals and you use n=26 years of data.

However, for the standard error of the regression, this formula should be replaced by



where k is the number of parameters in your regression model.

For a simple linear regression, k=2  because there are 2 parameters: the slope and intercept of the regression line.

In your case, k=4  because your regression model has 4 parameters: one parameter for each of your 3 predictor variables (area, extent and snow cover) and one intercept parameter.

That gives me a regression standard error of 0.38 M km2 for your model.

[Peter Ellis]

Here is what this hind-cast method did for the past 26 years :

Presumably at least some of those years were the ones used to fit the regression line, rather than being true hind-cast years.  Which of those years were not part of the data set used to fit the regression?

[cesium62]

it is highly unlikely (less than 2.5% chance) that Sept 2018 will end up below 4.5 M km².

You are saying that there is a <2.5% chance of this year's final outcome being about half a grid line below the prediction.  There are a couple of other years (2006, 2001, maybe 2010 and 2016) and  where the final outcome is about half a grid line away from the prediction.  It doesn't seem right that 10% of the years should display an event that should occur less than 2.5% of the time...

[Rob Dekker]

As a result, prediction for Sept 2018 September sea ice extent is quite high at 5.19 km², with a standard deviation of 340 k km².

I tried to reproduce your calculation, using the Rutgers and NSIDC data for June.  I reproduced your prediction of 5.19 M km2 for the September 2018 NSIDC extent.  But for the standard error of the regression, I calculated 0.38 M km2 rather than 0.34 M km2.

Thank you Steven, for taking the time to implement this method, and I'm glad you could reproduce the 5.19 M km2 prediction. That means that I explained the method well enough, and didn't make any mistakes in numbers or calculations (which is always a worry of anyone working with climate data).

Also thank you for correcting the standard error of the regression.
You are right, in that I calculated the SD using the simple formula, and I should include the extra variables. But even 0.38 M km2 is still quite respectable, no ?

[Rob Dekker]

Presumably at least some of those years were the ones used to fit the regression line, rather than being true hind-cast years.  Which of those years were not part of the data set used to fit the regression?

Normally for a hind-cast, that would include ALL the prior years available.
But I stopped doing that in 2016, since each year the standard deviation (and prior year predictions) change when you update with the next year observation.
To avoid that, and also to see how well my method would work in forecast mode, since 2016 I stuck with the 1992 - 2015 regression period.
So to answer your question directly : 2016 and 2017 were not used to fit the regression.

[Rob Dekker]

it is highly unlikely (less than 2.5% chance) that Sept 2018 will end up below 4.5 M km².

You are saying that there is a <2.5% chance of this year's final outcome being about half a grid line below the prediction.  There are a couple of other years (2006, 2001, maybe 2010 and 2016) and  where the final outcome is about half a grid line away from the prediction.  It doesn't seem right that 10% of the years should display an event that should occur less than 2.5% of the time...

Cesium, thanks for your comment.
I don't have the full series available on this computer, and Steven just made a valid comment regarding the standard error of the regression being larger than the 340 I presented.

So let me get back to your comment when I have the data available (which will be on Monday), but IIRC, only 2006 is outside of the 95% confidence interval of +/- 680 k km2 that I used.

[Peter Ellis]

Normally for a hind-cast, that would include ALL the prior years available.

No.  To be a hind-cast you have to use the algorithm to predict data points that ARE NOT PART of the dataset used for the regression.

Forecasting is predicting future data.  Hindcasting is "predicting" data that already exists, which you do by not looking at it before you make your prediction. That graph contains three forecast points (2016-2018) and no hindcast points.

[cesium62]

Rob: I think your model is fun and interesting, so I spent some time with it.

Assuming I've correctly re-produced your results...

The question above is, roughly:  what's the probability that the actual result is 0.7M km2 or more away from the predicted result?  This occurred in: 1980, 1982, 1983, 1986, 1988, 1991, 2001, and 2006.  That's 8 out of 39 years, or about a 20% chance.

Eyeballing the graph you published pretty much agrees that both 2001 and 2006 predictions are off by around 0.7M km2.


We can slightly simplify your approach by noting that you are predicting minimum extent as a linear combination of three variables: June snow area, June ice area, and June ice extent.  The multi-variable linear regression package that I'm using (XL miner in google sheets) notes that the 'extent' parameter isn't very useful in this prediction.  The software suggests there's a 6% probability that 'extent' should really be part of the equation.

Also, graphing the trend lines through the minimum and the predicted minimum, suggests that the prediction is diverging from actual (getting larger) as each year passes.  Since both snow cover and minimum extent trend downward year by year, it might be interesting to add 'year' as a parameter to better explore how well snow cover helps explain minimum extent.

Overfitting a model based on year, snow area, and ice area to all data from 1979 through 2017, we get the second attached picture.  And a forecast of 4.58 M km2 for the 2018 min extent.  (With a 360 K km2 geometric mean error.)  (Overfitting Dekker's model gives a forecast of 4.76 M km2 with a 435 K km2 geometric mean error.)  (If I train the year-based model on just 1992 through 2015, the forecast is 4.64 M km2 with a 386 K km2 geometric mean error.)

My simple physical explanation for the year-based model would be: heat is accumulating worldwide year by year due to greenhouse gases; the snow and ice area (or lack thereof) takes into account how much insolation is absorbed in the northern hemisphere in June.  Together, this suggests the amount of heat available for melting ice, subject to the vagaries of weather.

[cesium62]

Normally for a hind-cast, that would include ALL the prior years available.

No.  To be a hind-cast you have to use the algorithm to predict data points that ARE NOT PART of the dataset used for the regression.

Forecasting is predicting future data.  Hindcasting is "predicting" data that already exists, which you do by not looking at it before you make your prediction. That graph contains three forecast points (2016-2018) and no hindcast points.

Note that I have now provided a hindcast for the years 1979 to 1991 for Dekker's model in one of my jpgs above...

[Steven]

Normally for a hind-cast, that would include ALL the prior years available.

No.  To be a hind-cast you have to use the algorithm to predict data points that ARE NOT PART of the dataset used for the regression.

Forecasting is predicting future data.  Hindcasting is "predicting" data that already exists, which you do by not looking at it before you make your prediction. That graph contains three forecast points (2016-2018) and no hindcast points.

Rob used the words hindcast and forecast in exactly the same way as in the Schroeder et al. 2014 melt pond paper, e.g. see Figure 4 in that paper.  If you think their terminology is wrong, perhaps you should contact them about it.

Anyway, I made a variant of Rob's graph, in which I included the simulated predictions (blue color) that his model would have generated in previous years, using only information that was available at that time.  For example, to simulate a prediction for the year 2010, I used only data for 1992-2009 to calculate the coefficients of the regression equation. 

In contrast, the green points in the graph were all fitted from the fixed 1992-2017 regression equation and so they implicitly have knowledge that wasn't available back then.

As expected, the blue points perform somewhat worse than the green ones.



Here is a similar graph, using data from 1979 onwards


Edit: corrected graph

[Shared Humanity]

I think it is great that others are digging into data triggered by Rob Dekkers model but, if you change the graph, you should also change the title.