Automatically generated forecasts

A set of different automated systems are participating in the wetterturnier as well. These will be divided into the three groups “automated Systems” (MOS), “reference tips”, and “mean tips”. They are highlighted by colors in the ranking tables: human players [white], automated systems [black], reference tips [blue], and mean tips [orange].

  • automated systems: a majority of these are so called model output statistics (MOS) systems: statistically post-processed forecasts from a numerical weather forecast model
  • reference tips: special tips like e.g., persistence, or others (see below).
  • mean tips: there will be one mean tip including all submitted tips, and several group tips. The equation behind both is exactly the same. The only difference is, that the group tips are only including all active members of a specific group for a specific tournament


“Moses” is an automatic forecast based on the “model output statistics” (MOS) method. Carsten Raymund wrote his bachelor thesis in the year 2017 about this method. The bachelor thesis with the title “Regressionsanalysen zur Ableitung einer Optimal-Vorhersage, basierend auf der Kombination von Kurzfristvorhersagen der Teilnehmer am Berliner Wetterturnier” can be downloaded here [click]. Moses started its participation on 5.1.2001 with the equation following:

  • Moses = 0.5 * Dummschwätzer + 0.25 * Gatzen + 0.14 * Schmelzi + 0.11 * Frank

These coefficients are computed once a week with the results from the latest rounds/tournaments. If one of the users included in the equation does not participate it will be replaced by Petros. Worst case (if all players do not take part) Moses would be identical to Petrus. The current Moses equations can be found here:


The persistency forecast (“all parameters remain unchanged”) participates as a reference, and for statistical tests. We use the following observations:

  • Minimum-temperature (TTn) of Friday 6 UTC
    Cloud cover (N), wind direction (dd), wind speed (ff), weather state forenoon (Wv), atmospheric pressure (PPP), and dew point temperature (TTd) of Friday, 12 UTC
  • Weather state afternoon (Wn), and maximum temperature (TTm) of Friday, 18 UTC
  • Amount of Precipitation (RR) of Saturday, 6 UTC.


“Petrus” is the mean forecast of all participating contestants and methods of the weekend.

  • These parameters are averaged arithmetically: N, ff, PPP, TTm, TTn, TTd
  • For wind direction (dd) an average vector ist computed.
  • For precipitation amount (RR) a decision is made between dry cases and cases with observed precipitaion. Decisive is the majority of cases. In case of precipitation the arithmetic mean of all participants who forecasted measurable precipitaion is computed for RR. In case that the majority of forecasts are “dry”, the forecast of Petrus will be RR = -3).
  • The weather conditions (Wv, Wn) are decided by majorities. The following decisions are taken in a row (in case of no majority the higher code number is taken):
  • Decision dry / not dry weather condtions (0, 4 vs. 5, 6, 7, 8, 9)
    In first case simple decision between 0 and 4.
    In second case decision stratiform vs. convective (5, 6, 7 vs. 8, 9)
    In case of stratiform precipitation decision liquid vs. snow (5, 6 vs. 7), then decision between 5 and 6.
    In case of convective precipitation simple decision between 8 and 9.


Sleepy is no player. Sleepy defines the points a user gets in the ranking (total ranking) for the weekends where he has not participated in the tournament. The points Sleepy gets for a full weekend are mean points minus the ‘mean absolute residual’ to the mean points of all players. If p_c are the user-points for a given weekend and a given city c (max 200 points), and N the number of players for a given city, Sleepy (for city c) gets:

p_{sleepy,c} = \bar{p}_c - mean( | p_c - \bar{p}_c | ),~~\text{with}~~\bar{p}_c = \frac{1}{N} \cdot \sum_{i=1}^N p_i