20 Oct 2014 - Updated CV
Quick Links: My CV | AO Forecast | ENSO Forecast | MJO Forecast | Weekly CFS 2m Temps | Tropical Hovmöllers | Archived Hovmöllers | ENSO
Jul 112013
 

I have fixed some odds and ends over the past couple of days. Below is a list of the fixes.

1. Fixed the CFS Temperature Plots. There had been an error downloading the data over the 4th of July and until yesterday I didn’t have time to fix it.

2. Speaking of the CFS Temperature Plots, I have changed the variable displayed. They used to display the intraseasonally filtered forecast and spread. Now, they display the raw forecast and spread. The intraseasonally filtered forecast is a red line overlayed on the graph which looks a lot like a smoothing function. I think this approach gives more useful information.

3. The calculations of the spread anomalies were a little bit off before. I have fixed them. The error occurred in the calculation of the seasonal cycle of standard deviation of spread (try saying that 10 times fast!). The old procedure yielded values that were too small. Since this variable is the denominator of a fraction, the small values made the spread appear more anomalous than it actually was. Things should be much better now, though.

More fixes coming in the next few days… stay tuned…

 Posted by at 10:36 am
Jun 102013
 

Although we shouldn’t take the very long range CFS forecasts too seriously, it can be fun to look at them every now and again. I’ve had a page up for a while which lets you see the CFS 2m Temperature forecast for a specific point. Over the weekend, I made a horizontal map version of this, located here: www.kylemacritchie.com/cfstemps.

This page has 45-day 2m Temp and 500 hPa Height CFS forecasts for the North America region. If it proves popular, I will expand the domain. The initial maps you see are the unfiltered maps. If you press the down arrow, you will see the intraseasonal (> 2 weeks) filtered maps which show the larger-scale, lower-frequency pattern changes and should be more reliable.

The temperatures are shown in standardized anomaly form (sigma units). They were created by subtracting the long term mean (1979-2010), and the first 4 harmonics of the seasonal cycle, and then dividing by the seasonal cycle of standard deviation.

Note that the 2m temperature field is 0.5 degree but the 500 hPa height field is 2.5 degree resolution. Please let me know your thoughts, I’d be happy to tweak things based on demand!

 Posted by at 10:48 am
May 122013
 

Over the weekend I made some substantial changes to my GEFS maps. They are listed below in list form for your convenience.

1. The maps are now much bigger.

2. The entire page has been re-designed, hopefully it’s easier to use. There is a new menu navigation system. You can still get to it by going to: www.KyleMacRitchie.com/gefs . Please send me an e-mail if you have comments or questions!

3. I made the climo line on the spaghetti plots darker. Hopefully it helps.

4. New height anomalies! Instead of only shading spread, you can now view shaded height anomalies from climatology. Standardized versions of these maps will be available very soon.

5. Projection changes: Indian Ocean view expanded below Equator and North Atlantic view has been zoomed out and renamed to just “Atlantic”.

Information about the spread maps (From old post)

I’ve made some more tweaks to the GFS ensembles. Firstly, I’ve added an “Asia” domain to help out some viewers from that region. Secondly, I’ve added a new type of standardized anomalies that use sigma values instead of percentiles.

The new anomalies are created by assuming that the ensemble spread data follows a lognormal distribution. This isn’t really as fancy as it seems, it just means that when one takes the log of the data, it appears more normally distributed. The standardized anomalies are created by:

1. Obtain ensemble spread for each forecast time for each day from 1985 through 2010.

2. Take the natural log of these data.

3. Calculate the long term mean and first four harmonics of the seasonal cycle for each gridpoint at each forecast time over the climatology.

4. Use the climo data from (3) to create the standardized anomalies using typical z-score techniques: z = log(ens_spread) – (mean + seasonal cycle) / standard_deviation(seasonal cycle).

This allows us to take into account the fact that ensemble spread changes quite a bit throughout the year. It is important to note that these new maps don’t necessarily show new information that is not on the percentile maps, rather they show the same thing in a different way. It’s a matter of personal preference, but I think a lot of us, myself included, are used to looking at sigma anomalies over percentiles.

Because these are approximately normally distributed, you can apply your familiar 68-95-99.7 rules from statistics.

For each tail the corresponding percentages are, approximately:

1 sigma = 15.8% of the data lie above/below this.

2 sigma = 2.2% of the data lie above/below this.

3 sigma = 0.1% of the date lie above/below this.

Hopefully you can all choose your poison! I’ve updated the /gefs link to now go to the sigma plots, but the percentile plots will still be around and updated at the same time. Enjoy!

Information about Spaghetti Plots (from old post)
I have updated the GEFS spaghetti plots to show more information. There is now a solid black line, labeled, that shows the ensemble mean position for each of the three contour levels. There is also a colored line, labeled, which shows the climatological position of the contour level (Note: this is accurate to within +/- 12 hours since my climatology is daily. Luckily, height climos don’t change much in 12 hours.).

 Posted by at 1:26 pm
Apr 142013
 

Effective as of the 4:30 p.m. update on Sunday, April 14:

I have updated the GEFS spaghetti plots to show more information. There is now a solid black line, labeled, that shows the ensemble mean position for each of the three contour levels. There is also a colored line, labeled, which shows the climatological position of the contour level (Note: this is accurate to within +/- 12 hours since my climatology is daily. Luckily, height climos don’t change much in 12 hours.).

You can check out the new maps by clicking here.

 Posted by at 1:25 pm
Apr 072013
 

I’ve made some more tweaks to the GFS ensembles. Firstly, I’ve added an “Asia” domain to help out some viewers from that region. Secondly, I’ve added a new type of standardized anomalies that use sigma values instead of percentiles.

The new anomalies are created by assuming that the ensemble spread data follows a lognormal distribution. This isn’t really as fancy as it seems, it just means that when one takes the log of the data, it appears more normally distributed. The standardized anomalies are created by:

1. Obtain ensemble spread for each forecast time for each day from 1985 through 2010.

2. Take the natural log of these data.

3. Calculate the long term mean and first four harmonics of the seasonal cycle for each gridpoint at each forecast time over the climatology.

4. Use the climo data from (3) to create the standardized anomalies using typical z-score techniques: z = log(ens_spread) – (mean + seasonal cycle) / standard_deviation(seasonal cycle).

This allows us to take into account the fact that ensemble spread changes quite a bit throughout the year. It is important to note that these new maps don’t necessarily show new information that is not on the percentile maps, rather they show the same thing in a different way. It’s a matter of personal preference, but I think a lot of us, myself included, are used to looking at sigma anomalies over percentiles.

Because these are approximately normally distributed, you can apply your familiar 68-95-99.7 rules from statistics.

For each tail the corresponding percentages are, approximately:

1 sigma = 15.8% of the data lie above/below this.

2 sigma = 2.2% of the data lie above/below this.

3 sigma = 0.1% of the date lie above/below this.

Hopefully you can all choose your poison! I’ve updated the /gefs link to now go to the sigma plots, but the percentile plots will still be around and updated at the same time. Enjoy!

Click here to check them out!

 Posted by at 12:27 pm
Mar 222013
 

Some of you may have noticed some issues with the real-time images lately. Below is an explanation of the update times of the different types of images, along with details on their recent problems.

GEFS Maps
These maps are updated 4 times per day. The GEFS lags behind the GFS deterministic, so don’t expect updates at the same times. Unless there are server problems, these maps should be updated at 1z, 8z, 14z, and 20z each day. Recent problems have prevented this from happening, but I think that they are fixed now.

Teleconnection Indices
I recently made some changes to the teleconnection images, including removing the ensemble spread graph, which was bulky and unused, I think. I also tried to make the image format JPEG, but I encountered a server-side issue with Image Magick’s “Convert” command that prevented the images from being made each morning since the update. Instead, I have increased the quality of the PNG files that were generated, and hopefully those will now update on schedule. They are usually updated each morning by 9z.

I’ve also added 4 ensemble members to the teleconnection (and CFSR temperature forecasts) images using a “poor man’s ensemble” technique – yesterday’s 4 members are added to today’s for a total of 8. This will, hopefully, help the ensemble stay a bit more consistent from run to run. I argue that this works well because the low-frequency component of the circulation, which these teleconnection patterns target, should not change much from day-to-day.

 Posted by at 5:14 pm
Mar 182013
 

I’ve been working on some fun new graphs in an effort to make use of the GEFS and CFS forecasts as best as possible. I have created forecasts of 2m temperature from the GEFS and CFS forecasts all around the world. All temperature forecasts show anomalies from the seasonal cycle, calculated by removing the mean and first 4 harmonics.

When you click a point, the page will generate a graph for the closest grid point to where you clicked, on the fly. It may take up to 20 seconds if the server is very busy. Please be patient.

Explanation for Long Range (CFS, 0.5 degree horizontal resolution)
Each plot shows the CFS 2m Temperature forecast for 60 days. It has been filtered to remove high frequency variability (< 2 weeks). This is important because the high frequency variability in a long-range forecast is unlikely to be consistent from run to run and isn’t much more useful than noise. However, the lower-frequency (intraseasonal) signal is more consistent, and more useful in the long-term.

The solid black line is the ensemble mean and the shading shows the extent of the ensemble members. Keep in mind that the CFS only has 4 ensemble members and they aren’t made as cleverly as the GEFS ensemble members, so they don’t always behave very well.

You can view the forecasts here: http://www.atmos.albany.edu/student/macritch/showcfs2mplots.php

Please remember that these are in beta, so they will probably have some issues. Please let me know if you find any!

*SHORT RANGE IS COMING SOON*
Explanation for Short Range (GEFS, 1 degree horizontal resolution)
Each plot shows the GEFS 2m Temperature forecast for 15 days. The GEFS forecast shows standardized ensemble spread, in percentiles, so that you have an idea of how confident the model is in the forecast. The standardized ensemble spread is explained in the previous post.

In general, high percentiles indicate anomalously low confidence and vice-versa. Keep in mind, though, that while low ensemble spread always indicates a lack of confidence, high ensemble spread only usually indicates a lack of confidence.

 Posted by at 8:12 pm
Feb 042013
 

I’ve added some new GFS Ensemble real-time maps. I now have heights at 200, 250, 500, 700, and 850 hPa. Plotted on these maps are the ensemble mean and standardized ensemble spread.

You can also view the GFS ensemble maps with all the levels by clicking here. Just use your up and down arrows to go through the atmosphere!

What is standardized ensemble spread?

I used the GFS 2nd generation reforecast data to create a climatology from 1985-2010 of ensemble spread at each forecast time. Since ensemble spread is not normally distributed, I couldn’t use the typical sigma values that we’re used to with anomalies. So, I opted to show percentiles. For those who need a little refresher – percentiles go from 0% – 100%, I’ve calculated them in 5% intervals but I only plot 10% intervals. Low values imply anomalous predictability, high values imply an anomalous lack of predictability and middle values (~50%) imply normal values. A percentile value of 80%, for instance, means that 80% of the values in the climatology fall below this value, or that 20% of the values lie above it.

What’s the point of standardized ensemble spread?

Two major reasons come to mind.

1) Regular ensemble spread is meaningless if you don’t know where it lies in the climatology. The reason for looking at ensemble spread is to get a rough idea of the confidence of the forecast, but without standardized values, the forecast confidence is difficult to determine.

2) Regular ensemble spread is not comparable among forecast times. 100 meters of ensemble spread in a 12 hour forecast is very different than 100 meters of ensemble spread in a 6 day forecast. Because the standardized ensemble spread is normalized at each forecast time, it is possible to compare spread among different forecast times.

Caveats

1) In general, high ensemble spread implies a lack of predictability and low ensemble spread implies increased predictability. There are, however, exceptions to this rule which will be discussed in an upcoming post on ensemble forecasting.

2) The GFS 2nd generation forecast uses a frozen model of the GFS from early 2011. This is not the current version. This makes the climatology slightly incompatible, but it’s the best we’ll find. Instead of comparing apples to apples, think of it as comparing granny smith apples to red delicious apples; it’s still much better than comparing apples to oranges.

Give the GEFS maps a try and let me know if you like them! I’m open to any suggestions!

 Posted by at 10:34 am