Yandex.Meteum - technology without technology. Marketing accurate to the district

    It has been exactly 3 years since the launch of the Yandex.Meteum service, which, according to the developers, gives the highest quality forecasts among all the forecast sites. It's time to take stock. Is the new product of Yandex really revolutionary or is it all just marketing stuff? And as a team of scientists from the Hydrometeorological Center of Russia, we managed to bypass Yandex and create a truly high-quality forecast.

    For a start I will introduce myself. My name is Ilya Vinshtein, I am an amateur weather forecaster from Kurgan. I am engaged in meteorology for 14 years. I administer my regional project “Weather 45” , gave several lectures for the scientific and educational project “The Smoking Gutenberg”.

    Meteum. Beginning The

    problem of the quality of forecasts of the Yandex service. Meteum ”I identified at the beginning of 2016 , but at that time I had a very small amount of data to make unambiguous conclusions. I mainly criticized the information campaign that Yandex launched. If at the very beginning they used the slogan “Forecast with accuracy to the house”, then they changed it to “Forecast with accuracy to the region”.

    In what form did the birth of Meteum take place? There were a lot of publications in the media, several publications on Habré, then short tutorial videos went on to explain the creation of the first predictive service based on the neural network. Creation of the Meteum came at a time when neural networks were very popular, and the media presented them as a panacea that could solve many problems of humanity.

    In the wake of the general neuroauge in Yandex decided to update its main meteorological section. Regular updating of the design and adding new functions is not an option, and the main goal of the update is to attract new audiences from competitors. The only way was to change the paradigm of service perception. Not just a weather section, but a completely new service that can predict the weather better than all other competitors. Not just “Yandex. Weather, ”and“ Yandex. Meteum "- a system capable of issuing a forecast with an accuracy of a house or area.

    The problem is that the final product is difficult to evaluate and receive feedback. Who will check the quality of forecasts? How will the feedback from users be evaluated? In our case, the consumer is unable to assess the quality of the final product, so he can “vparit”, anything can be done. Users of all forecast sites use this trick.

    Therefore, the entire information campaign on the launch of Meteum looks like an anti-scientific farce. For all 3 years, Yandex has not provided us with any objective figures. We have not seen a single report on the justification.

    All numbers come down to this:
    “According to our own estimates (alas, there are no independent meters in this area yet), for today our weather forecast is more accurate than all the competitors we know. For example, the temperature forecast for 24 hours is mistaken for us by 35% less than the nearest competitor. ”
    This is an absolute lie. I will explain why. If the year was 1960, then this statement was absolutely fair, but now the short-term forecasts have already reached a certain ceiling. The struggle goes for interest and even a few tenths of a percent. For example, according to the Hydrometeorological Center of Russia, the accuracy of the forecast of air temperature in Russia in 2017 for a day was 93%. We are talking about those forecasts that were given by weather forecasters of local weather centers. For predictive sites and computer models, the accuracy varies from 85% to 95%. No 35% here and there can not be!

    How does this happen in the world of science?

    Suppose that Yandex managed to create a revolutionary product that truly bypasses all competitors. Introduce the world to this technology. Let the Yandex team show the world what they managed to create. Why not start with an article in a scientific journal? I understand that maybe it is not necessary to disclose all the cards, because the product, in fact, has a commercial component, there is nothing wrong with that, but only if this product really works.

    Any scientific forecasting method passes the testing phase. Usually experimental hares are archival data. In some cases, it is necessary to collect data for a year or several years, and only then publish the article. After this, the forecast is compared with the actual weather station reports. Prognostic fields are related to actual ones. The mass of parameters is calculated: the arithmetic mean temperature error, the average absolute temperature error, the relative error and justification in percent. Then, based on the results of operational tests, a special methodical commission makes a decision to recommend or refuse the use of this forecast method.

    And now a question for the developers of Yandex. Where is this data? Where are these articles and studies? "We have the most accurate forecasts, believe us," say in Yandex. Nothing to show. Nothing to brag about.

    We have the numbers, but we will not show them to you. You have no documents.

    Ensemble and multi-model forecasts

    Recently, ensemble, multi-model and complex forecasts have won great popularity among weather forecasters. What does it mean? First, a little theory. The main source of all forecasts are computer models. Programs that simulate the entire atmosphere of the Earth, starting from the soil and ending with the upper layers of the stratosphere. The main food for the models are satellite datain all visible and invisible spectral ranges. Data of ground stations now do not have such a strong influence on the quality of the forecast. If we exclude a layer of weather stations from models, the quality will fall by 7%, and if we exclude satellite data, then by 35-40%. In the world there are 11 global models and a dozen more regional.

    Computer models are very, very complex! Not every state is able to create its own qualitative model. For example, the domestic model PLAV occupies the 8th position in the ranking of world models. It exists, but almost never used.

    For this reason, most sites and applications use only 2-3 models. Everything else is a matter of internal data processing and interpretation. For example, now the European Medium-Term Weather Prediction Model (ECMWF) is the best model. This model is used by Foreca, intellicast and Gismeteo. BUT! As I said, the processing of forecasts is reduced to internal patterns that are engaged in "grinding" the raw files of the model. Gismeteo makes it worst of all, and Intelicast makes it better. Next will be the numbers confirming this.

    Okay, you figure it out. One run of a computer model is in its pure form a deterministic forecast on an “as is” basis. The main problem of deterministic forecasts is errors in the initial data, which lead to the butterfly effect. The smallest initial disturbances lead to huge errors in the medium term. To solve this problem, scientists have developed ensemble forecasts . Imagine the usual deterministic forecast. An artificial error is introduced into this prediction using a pseudo-random number generator.

    Ensemble forecast. American model GFS. 20 members.

    And so it is done another 20 or 50 times. Then a graph is drawn up, where you can see how sensitive the forecast is to errors in the initial data. If the deterministic forecast gives a warming after 10 days, and 20 ensemble members go down, that is, they give a cold snap, then the deterministic forecast for this period is wrong.

    But scientists went even further. They began to synchronize deterministic forecasts and create multi-model forecasts , when the forecast is not built on the basis of one model, but just a dozen.

    Multi-model forecast for Moscow on the site meteoblue. 11 models

    For example, 7 models give precipitation after 5 days, and 3 predict dry weather. Therefore, the probability of precipitation is 70%. Together recommended to watch more and ensemble forecast.

    And now we are getting to the point. How did the weather center manage to bypass Yandex?

    Comprehensive forecast

    In 2014, the head of the hydrodynamic short-term forecasts of the Hydrometeorological Center of Russia and honored meteorologist Alexei Bagrov, together with his team, developed a simple but fundamentally new statistical scheme for processing raw forecast data . It was published in the journal "Meteorology and Hydrology" in an article entitled "Comprehensive forecast of surface meteorological variables."

    The essence of the technique is simple, but this is its superiority. Comprehensive forecast obtained by statistical processing of the results of the included models. At the same time, for the air temperature, wind and dew point, an archive of forecasts for the previous 20 days is used for the corresponding models and actual data at the station, and for precipitation a similar archive is for one year. The calculation is carried out separately for each station and for each forecast lead time.

    If it is even simpler, then Bagrov suggests performing a statistical adjustment of the forecasts of the best models based on the actual data of the local weather station. The method is described in detail in the article itself.. Here I will focus on some highlights. The calculation of the maximum and minimum temperature is performed taking into account the error for the last 5 or 3 days. For example, over the past 5 days, our models have lowered the temperature by an average of 2 degrees, so we need to include this error in the last forecast and stabilize the forecast to the most likely value. Thus, the forecast itself automatically adjusts itself, relying on previous deviations in the direction of overestimation or understatement.

    4 year forecast was in the testing stage. In September 2018, the test results were published in the journal Russian Meteorology and Hydrology . Briefly, the research results are announced here.. I will note that all 4 years the forecast was published on the website of the methodical cabinet of the Hydrometeorological Center of Russia. It was calculated for 224 cities of Russia. Every month was a report of acquittal. They continue to go to this day.

    Modestly and quietly - they created the best forecast.

    The Hydrometeorological Center of Russia did something that nobody else could do. They automatically collected forecasts from 7 different forecast sites and analyzed their accuracy. Below are data for 1.5 years - from January 2016 to June 2017 for Moscow, St. Petersburg and Yakutsk.

    The average absolute error of the forecasts of the minimum (a) and maximum (b) temperature by city: Moscow, St. Petersburg, Yakutsk for the period January 1, 2016 - June 30, 2017 Site forecasts: 1 -; 2 -; 3 - Fobos (; 4 -; 5 -; 6 -; 7 - Complex forecast of Bagrov.

    We got to the bottom. From the data it can be seen that, on the first day of the daytime temperature, Yandex bypasses 3 resources at once: meteoinfo, intellicast and Bagrov’s integrated forecast. The latter shows the lowest error for 1-2 days. For 3-4 days Intellicast and a comprehensive forecast lead. Yandex only 3 positions.

    Please note that the most popular Gismeteo in RuNet is not that precise. On the first day, its average error of 2 degrees is a lot. The ranking leader is the site

    No need to think that there is no more fresh data. At the beginning of 2018, a section entitled “Evaluating Forecasts on Various Internet Sites” appeared on the website of the methodological cabinet of the Hydrometeorological Center of Russia . The section publishes data for 47 cities both individually and together.

    Many may say that this is outdated data, but there is already a fresh report of justification for October. We study it. Take a sample of 27 cities for ETR.

    The forecast of daytime temperature at Yandex for one day is comparable to the accuracy of intellicast and a complex forecast. For the next 2-5 days, intellicast slightly bypasses Yandex. With the forecast of night temperature at Meteum everything is somewhat worse. On the first day, it goes around 3 sites: meteoinfo, intellicast and a comprehensive forecast. The next day the trend continues. On the 6th day Yandex overtakes intellicast and meteoinfo.

    For the Asian territory, the distribution is approximately the same. In almost all cases, Yandex bypasses the meteoinfo, intellicast and complex troika. Many have noticed that a good accuracy is provided by the official website of the Meteoinfo Hydrometeorology Center. Yes it is. Now the site uses an independent statistical scheme for processing model data called REP (calculation of weather elements). This scheme is not bad, but somewhat worse than a complex forecast. In the winter, she doesn’t predict good nightly cooling. I draw your attention that all these data processing schemes were invented long before the creation of the Yandex PR. Meteum.

    Output and Display Issues

    It is not enough to create a qualitative forecast, it is necessary to learn how to adequately display it for the average person. When a user visits a weather site, he first looks at the forecast for 10 days, getting the overall picture of the temperature change. But if you dig deeper, looking at the course of temperature, then many nuances will open up. For example, the site indicates that the day will be +15 degrees, but then you open the temperature course and understand that these +15 will be at night, and during the day the temperature will be lower! This situation is called reverse temperature when it is warmer at night than during the day. The problem of outputting the maximum and minimum temperatures here is that the min and max values ​​are captured from the entire time series, not dividing day and night. From the point of view of the average man - this is a hoax. This sin all sites. The maximum temperature is usually recorded from 08 to 20 hours, depending on the time of year, the weather situation and the coordinates of the weather station. The minimum is noted from 20 to 08 hours, again, depending on the season and synoptic situation. This is called a meteorological day. For example, the main weather stationMoscow to VDNH sends the maximum temperature at 21 o'clock, and the minimum at 9 o'clock.
    Below, I have brought situations of an atypical temperature variation, when it is important to capture the maximum and minimum temperature not from the entire time series, but at strictly fixed intervals. If the conditions are not met, the user will be deceived, even despite the qualitative forecast.

    Another problem is that a couple of years ago, Yandex began issuing climate data for a long-term forecast, which is not entirely correct. In Yandex, it was decided to use the raw CFSR (NCEP) computer reanalysis files for the last 7 years, creating a small climate sample. Now they have switched to the averaging period of 10 years, which does not change the situation. Also on the site appeared the parameter "Probability of precipitation", which was similarly calculated over the past 10 years based on computer reanalysis, but there is a serious problem. Computer reanalysis very poorly simulates convective rainfall and light snowfall in winter, so Yandex could issue a 0% probability, and according to the weather station it was 50%, only because there is a precipitation meter at the weather station that captures real precipitation, but not virtual . Therefore, it is more correct to average the data on the weather station, and not on computer reanalysis. I wrote to Yandex and received a response: “We have added it to the list of suggestions from our users. Our specialists are always acquainted with these ideas when they prepare any changes in the service, and try, if possible, to take them into account. ” After 6 months, nothing has changed.

    Staying alive (c)

    I also suggested that Yandex, instead of averaging data over 10 years, use forecasts of the long-term climate model CFSv2. It is updated 4 times a day and considers the forecast for 9 months ahead. Of course, we are talking about obtaining a decade or monthly average data. But this is a real forecast, not historical information. For example, now the model gives out that November in the European part of Russia will be warm and dry .

    It was especially amusing to watch how Yandex epically rolled out prognostic maps, although at that time,, and already existed. These services provide many times more information on various models. They wrote in Yandex that the main difference from them is that they have higher accuracy. Oh well.

    The problem of night cooling

    Now in synoptic meteorology, the question of predicting the night temperature in the conditions of anticyclonic air cooling is still an urgent issue. What is the problem? The problem is that computer models almost always overestimate the temperature in such a synoptic situation. For example, in Kurgan, according to most forecasts, at night the temperature will drop to -30 degrees: it will be clear, the pressure will rise, and an anticyclone core will pass over the region. Ideal conditions for cooling. But in fact, the minimum can drop to -35 ...- 37 degrees! When Yandex launched Meteum, I thought that finally it would be possible to solve this problem. But already that winter, Yandex continues to inflate the night temperature in the conditions of cooling. At the moment there are only 2 models in the world that can adequately calculate this temperature. The first is the Canadian GEM model. The second is the North American NAEFS. In other synoptic situations, these models do not show anything outstanding, so my task as a forecaster is to include these models in the forecast at the right time, and then at the right time to exclude them from the forecast. Even with the cooling cope complex forecast Bagrov, just due to statistical adjustments based on the local weather station.


    The data obtained allows us to say with confidence that the temperature forecast of Yandex using the Meteum technology does not show any exceptional results compared to the already well-established forecast services. On the contrary, Yandex is inferior in quality to such resources as intellicast, meteoinfo and Bagrov’s integrated forecast published on the website of the methodological office of the Hydrometeorological Center of Russia. The worst weather services can be called Gismeteo and

    There is science, and there is marketing. One goal is to achieve a result, in our case, high accuracy. In the other, the goal is to gain and receive an imaginary and bloated achievement that does not carry any practical significance for the average man. Yes, the Hydrometeorological Center is a state organization that does not have such advertising budgets and opportunities as Yandex has, but the main thing is that there are still people there who can not create money, but science.

    Also popular now: