Back - Trigger

Software Projects - Exploratory Data Analysis

Lacrimae rerum. Memento mori. Memento vivere.

Social Media Posting Times

The time at which a post is made is often seen to be important in determining the resulting performance. The recommendations often range from posting when no one else is posting (and fewer users are typically online) to posting when most users are online (and everyone else is typically posting). However, from a rational perspective, it should be expected for the influence from this time to be inconsequentially minimal relative to the other factors which affect performance. These other factors, such as quality, format, relevance, audience, and industry, should be much more important. This should especially be true with regard to long-term performance, although short-term performance many be impacted by idiosyncratic influences. An analysis was performed for investigation of these ideas.

Analysis Methodology

When investigating the effect posting time has on performance, it is necessary to consider the long-term influences, as well as the short-term influences resulting from posts published at different times and on different weekdays. With regard to the long-term influences, this can be analysed by simply using a performance metric at the end of the lifecycle of the post. With regard to the short-term influences, this is more tricky due to the frequency at which data is collected, but it can be analysed by considering the rate at which a performance metric increases in a given period after the post is published.

Essentially, it is necessary to isolate posting time and performance as the features for direct investigation. This requires normalization of the data to control for other features, such as the audience size, industry, growth, and other idiosyncratic factors related to the array of accounts. To do this, the posts of each account are individually considered within a group, where a standard score is calculated based on the performance relative to the historical performance of other posts from the account. The performance metric considered is the log of engagement (although it is also shown that impressions are correlated and will yield a similar result). The standard score is calculated as the ratio of the difference from the mean over the standard deviation and will provide a measurement of where the post was positioned in relation to other posts as expected/normal or an outlier. A moving window of 360 days is used for considering the historical performance of other posts from the account. The data discussed also spans over 3 years to mitigate any transient or algorithmic complications, such that the results should be objective regardless of momentary conditions.

With a standard score for each post, it is then possible to find the average standard score for the posts which fall into groups based on their time and weekday of publication. If there is a specific time and weekday which is better for publishing, then we would expect this group to consistently have a higher standard score on average, as these posts will consistently be outliers relative to posts published at other times and weekdays. If there is no specific time and weekday which is better for publishing, then we would expect for all of the groups to have a standard score approaching zero as the size of the sample is increased, as any outliers will be randomly distributed based on other features. The most crucial assumption is that the posts are sufficiently distributed throughout the times and weekdays within each account (as a variety is obviously required to be the case for any investigation into posting time).

[Got a figure in mind if we want to illustrate this. Illustration of the methodology and expected results for whether or not posting time is relevant].

For additional context, the log of engagement is considered for the performance metric, as the distributions of engagement appear to consistently follow a log-normal distribution for most accounts (as is somewhat expected since the performance metrics are bounded on the left at zero and infinite on the right), such that the log of engagement should be more normally distributed. Although the results were not fully tested apart from quantile-quantile plots, the distributions do appear to at least be more symmetric with the moderation of extreme outliers from skewing the distribution.

LinkedIn Long-Term Results

Considering the overall distribution of values for all of the accounts, there are qualitative similarities shared with a normal distribution. This includes symmetry around the mean with approximately equal values for the mean and median near 0 (specifically, these are equal to -0.01199 and -0.04458 respectively). The data also tends to be clustered around the mean, where there are fewer values in the tails moving away from the mean. However, a quantile-quantile plot was used to show how the distribution compares relative to various other distributions. With regard to a normal distribution, it is seen that there is a resemblance, although the points deviate from the diagonal line particularly at the tails - actual quantiles are generally lower than the corresponding quantiles of a normal distribution. This suggests that the data appears to have lighter tails than a normal distribution, which means there are less extreme values in the data than would be expected in a normal distribution. Still, for the nature of this investigation, a normal distribution is a reasonable assumption.

Distributions of the standard scores applied to the log of engagement for different accounts:
Overall distribution of the standard scores applied to the log of engagement across all of the accounts:
Quantile-quantile plot showing the line of best fit in comparison to a normal distribution as the theoretical target:

With the standard score for each post, the groups for posting time are created using the hour and day in which the post was published. The mean of the standard score is calculated for each group. In order to ensure that there are enough posts within a group, a threshold of 32 posts in the sample is applied for filtering (arbitrarily chosen to provide a clean view of results from Monday to Friday between 6:00 and 21:00).

Aggregation of standard scores of long-term performance using the hour and day of posting on LinkedIn:

From this aggregation, it is seen that the mean standard scores vary between 0.21 and -0.19. There is also no discernible pattern in the distribution of values between time and weekday. Thus, the results clearly indicate that the posting time does not have a significant influence on performance compared to other features. This should be expected, as the quality and interest of the content will almost always be the most important and consistent features in determining performance (while any abnormal performance as a consequence of "virality" is most likely to be random due to luck).

LinkedIn Short-Term Results

For short-term considerations and due to the frequency at which the data is collected, the engagement is divided by the time since publication relative to the time at which the data was collected. This is necessary for a fair comparison when the performance was recorded at different times. An additional filter is also applied to select the latest data within the first 24 hours after being published. Essentially, this leads to a linear rate representing the increase in engagement per unit time since publication. This assumption is valid if the change in engagement within the first 24 hours after being published is approximately linear (which is somewhat valid, although any time after this has been seen to decay exponentially).

Example of the linear approximation used for the first day after publication in which the performance is increasing:

As with the long-term considerations, the overall distribution of values for all of the accounts can be seen to share qualitative similarities with a normal distribution. This includes symmetry around the mean with approximately equal values for the mean and median near 0 (specifically, these are equal to -0.01120 and -0.03489 respectively). The data also tends to be clustered around the mean, where there are fewer values in the tails moving away from the mean. However, a quantile-quantile plot was used to show how the distribution compares relative to various other distributions. With regard to a normal distribution, it is seen that the points deviate from the diagonal line particularly at the tails - actual quantiles are generally lower than the corresponding quantiles of a normal distribution. This suggests that the data appears to have lighter tails than a normal distribution, which means there are less extreme values in the data than would be expected in a normal distribution. Still, for the nature of this investigation, a normal distribution is a reasonable assumption.

Distributions of the standard scores applied to the log of engagement per unit time for different accounts:
Overall distribution of the standard scores applied to the log of engagement per unit time across all of the accounts:
Quantile-quantile plot showing the line of best fit in comparison to a normal distribution as the theoretical target:

With the standard score for each post, the groups for posting time are created using the hour and day in which the post was published. The mean of the standard score is calculated for each group. In order to ensure that there are enough posts within a group, a threshold of 22 posts in the sample is applied for filtering (arbitrarily chosen to provide a clean view of results from Monday to Friday between 6:00 and 18:00). It should be noted that the amount of data is reduced compared to the long-term considerations due to historical inaccessibility.

Aggregation of standard scores of short-term performance using the hour and day of posting on LinkedIn:

From this aggregation, it is seen that the mean standard scores vary between 0.41 and -0.33. An explanation is proposed below, but there is no discernible pattern in the distribution of values between time. There may be a noticeable pattern in the distribution of values between weekdays, where Mondays perform better than Thursdays and this can be somewhat understood based on the nature of LinkedIn - it is likely that users stay off the network on weekends but logon on Monday to catch up what they may have missed, but they have already caught on any news or announcements and are focussing on other work by Thursday. Thus, the results clearly indicate that the posting time may have a slight influence on performance in the short-term but this is diminished over time. This should be expected, as the quality and interest of the content will almost always be the most important and consistent features in determining performance (while any abnormal performance as a consequence of "virality" is most likely to be random due to luck).

There may be a minor hint that posts published later in the day perform better than posts published earlier in the day. However, this may be related to the assumption that the change in engagement within the first 24 hours after being published is approximately linear. As seen this is not completely accurate, where posts usually have the steepest gradient closest to when they are published with a decline from that point. The data was collected each day at 3:00 in the morning, so posts published later in the day may have a very slight advantage relative to posts published earlier in the day due to the frequency at which the data was collected. There may be alternative reasons, but this should be kept in mind.

Example of the limitation which may favour posts published later in the day over posts published earlier in the day:

Other Network Results

Expanding the investigation, a similar analysis can be performed for Facebook, Instagram, and Twitter. However, in most cases, the amount of data available is more limited, so it is understandable to expect a greater dispersion in results (with completely unreliable results for the short-term analysis in some cases). Regardless, similar results are seen, where there is no discernible pattern in the distribution of values between time and weekday.

Overall distributions of the standard scores of the log of engagement for Facebook, Instagram, and Twitter:
Quantile-quantile plot showing the line of best fit against a normal distribution for Facebook, Instagram, and Twitter:
Aggregation of standard scores of long-term performance using the posting time on Facebook, Instagram, and Twitter:
Overall distributions of the standard scores of the log of engagement per unit time for Facebook, Instagram, and Twitter:
Quantile-quantile plot showing the line of best fit against a normal distribution for Facebook, Instagram, and Twitter:
Aggregation of standard scores of short-term performance using the posting time on Facebook, Instagram, and Twitter:

Summarized Conclusions

As a final note, these results should be completely expected if social media behaves efficiently. For example, if there was a special posting time for dramatically improved performance, everyone would start to post at this time which would consequently saturate and eliminate any perception of improved performance. In conjunction, the algorithms developed to serve content are extremely complex with many variables for consideration, so it is somewhat unreasonable to expect something as simple as posting time to have a noticeable weight in deciding whether content is relevant. Thus, on average, the posting time is never going to have a dramatic long-term effect on performance.