Problem: It is known that weather can affect mood, but until recently, little was known about the details of this association.
Solution: Valuable insights can be gained from social network posts reflecting users’ moods at the time of posting.
Sergey Smetanin, Research Fellow of the HSE Graduate School of Business, conducted a large-scale analysis to examine the impact of weather conditions on the sentiments expressed by users of the Odnoklassniki (OK) social network. He reviewed a total of 2.76 million posts published by 1.31 million unique users and found that the highest number of posts expressing positive sentiments were associated with maximum average daily temperatures between +20°C and +25°C, a light breeze (5 to 11 km/h), and an unusual increase in temperature (20°C or more) from the previous day (eg from -5°C to +15°C). The findings have been published in PeerJ Computer Science. This is the first study of its kind in Russia.
Science has shown that weather can clearly affect people's behaviour and mood. However, prior studies of the association between weather and mood have produced mixed results. This, Smetanin argues in his paper, was partly due to the lack of big data on emotional states. One feasible solution is to examine the digital traces left by social media users, who often indicate their current mood through hashtags on their posts and by other means.
For example, a paper based on two years of user data from Twitter (access to Twitter is blocked in Russia quoting 'violation of the key principles of free dissemination of information and unhindered access for Russian users to Russian media on foreign internet platforms') showed a correlation between Twitter users' emotional states and daily fluctuations in temperature, precipitation and snow depth. A study based on 3.5 billion posts from Twitter and Facebook (owned by Meta, recognised as an extremist organisation in Russia) found cold temperatures, hot temperatures, precipitation, narrower daily temperature ranges, humidity, and cloud cover associated with worsened expressions of user sentiment. Another recent study of more than one billion tweets found consistent and statistically significant declines in expressed sentiment from both hot and cold temperatures in the US.
In Russia, most studies of the relationship between mood and weather were limited to just one region of the country and had a sample of up to 1,000 people, and none examined digital traces. Smetanin’s work has now filled this gap.
Smetanin used data from the Odnoklassniki social network for his analysis. He observes that prior studies mainly relied on platforms such as Twitter, which may not be feasible in the Russian context, because it is not as widely popular with Russian users and open-access databases contain few Russian-language tweets with geotags.
According to the author, Odnoklassniki (OK) was chosen as one of the largest and most popular social networks in Russia and a low-threshold platform used by people of different ages and various levels of digital skills.
The dataset was obtained from OK’s Data Science Lab. The study analysed 2.76 million posts from 1.31 million unique users from seven Russian cities: Moscow, Yekaterinburg, Krasnodar, Novosibirsk, St Petersburg, Samara, and Rostov-on-Don, dated between March 2019 to March 2021 (two years’ worth of data).
The data was processed using RuRoBERTA-Large-RuSentiment, a machine-learning model for natural language processing, and categorised according to the mood expressed into five sentiment classes: Positive, Neutral, Negative, Speech Act (greeting, congratulations, etc), and Skip (mixed or difficult to interpret).
The historical weather data was obtained via Meteostat, an open database providing detailed weather information from thousands of weather stations and locations worldwide. The final analysis included data on wind speed, the difference in average temperature from the previous day, and temperature change during the day. Lastly, regression modelling was applied to investigate the relationship between weather and the sentiments expressed in user posts.
It was found that a light breeze and maximum daily temperatures between +20°C and +25°C without major fluctuations throughout the day were associated with the highest number of posts expressing positive sentiment.
This finding is consistent with studies in other countries—except that the latter demonstrate a decline in positive expressions past +30°C which does not correlate with the Russian data. While no such association was found in the current study, the reason may be that the Russian sample included very few days with average temperatures above +25°C, according to Smetanin.
In addition, Russian users were found to prefer a light breeze with wind speeds between 5 and 11 km/h but not between 11 and 19 km/h. Smetanin observes that this finding is consistent with those from earlier foreign studies. Other wind speeds did not demonstrate a statistically significant association with expressed sentiment in the Russian study.
A 15–20°C difference between the maximum and minimum daily temperatures was associated with fewer positive posts—in other words, Russian users are not likely to experience a positive emotional state when temperatures change drastically within one day. The baseline temperature range for comparison was set at 0°C (0°C < TTR ≤ 5°C).
No statistically significant association between expressed mood and temperature ranges below 15–20°C was found, although it would be logical to expect negative moods at lower temperature ranges. There may be several reasons why this was not the case. 'First, OK users tend to share a lot more positive posts than negative ones, so perhaps there wasn't enough data in the sample to find an association between weather and negative sentiment for the period in question. Second, certain weather conditions are less common than others (eg storm winds are rare in Russia), therefore available data on certain types of weather may have been insufficient for statistically significant results. Third, in my previous study, I compared expressions of negative sentiment in OK posts against relevant data from VCIOM surveys and did not find any statistically significant associations,' explains Smetanin.
Generally, the study findings suggest that people prefer gradual, rather than drastic, temperature change. This finding is not entirely consistent with previous studies. For example, one study reported an association between daily temperature ranges exceeding 15°C and a significant increase in positive expressions, while another study did not find daily temperature range to have any significant effect on sentiment.
Meanwhile, a substantial increase in daily average temperature by 20–25°C compared to the previous day is associated with more positive posts from OK users. 'In other words, when it suddenly gets warmer by 20 degrees, eg from minus 5°С yesterday to plus 15°С today, we observe a higher proportion of user posts expressing positive sentiment,' the researcher comments. This finding, once again, is not entirely consistent with earlier studies reporting that people tend to feel uncomfortable with drastic temperature changes from one day to the next.
Smetanin explains that this particular difference in findings is expected and, in fact, consistent with the conclusion of many other studies in this field that the way weather affects mood may be modified by additional factors such as local climate zone, economic and social conditions, individual differences, and seasonality.
Due to climate diversity across Russian regions, findings from foreign studies are unlikely to provide an accurate insight into the effects of weather on Russian people's emotional states. Understanding these effects can be important for studies in social science, human psychology, and medicine, the author notes.
It seems obvious that weather can affect mood, but research focusing on this relationship often produces mixed results. Over the past decade, the spread of social media use and the advancement of natural language processing have made it possible to study the relationship between weather conditions and internet users' moods. It is clearly important for businesses, when designing special offers and promotions, to understand and take into account changes in consumer behaviour caused by external factors.
The study has some limitations. For example, further research is recommended to improve the psychometric validity of sentiment measures based on automated processing of digital traces. In addition, there may be a discrepancy between weather parameters obtained through Meteostat and users' subjective experience of weather. The relationship between weather and expressed sentiment may also depend on current social and economic conditions. Therefore, further research in this area could be beneficial.
Academic Supervisor of the project, Professor of the HSE Graduate School of Business Department of Business Informatics