Daily iPhone App Sales Trends

<em>William Volk, CCO</em>

<em>Post Featured in <a href=”http://www.gamasutra.com/blogs/WilliamVolk/20110418/7458/Daily_iPhone_App_Sales_Trends.php”>Gamasutra</a></em&gt;

In marketing iPhone apps it is useful to see the effects of app description, screen shots, advertising campaigns and public relations on the sales of your app. The best way to do that is to see how these changes effect the number of downloads you are seeing on a daily basis.

There is a problem with this approach. What developers have long observed is that download volume varies, based on the day of the week. For example: If you changed a screen shot in your app description (on the app store) on Friday, is the 10% increase in sales on Saturday (compared to Friday) a good or bad thing or just a reflection of more downloads on Saturday vs Friday?

You could always compare the sales numbers to last week, but that could reflect larger trends. If you are making frequent changes it’s good to have a way to compare day to day sales/downloads. So we set out to determine the day to day trends for app downloads.

We started with a free popular game that had been released in January, Bocce-Ball. Bocce-Ball had hit #6 in all apps, and was continuing to see a good number of downloads even after promotions and advertising had stopped. What’s more, we were weeks away from an update so we believed we would see consistent download numbers and definite patterns emerging on the sales based on the day of the week. The weekly download volume remained relatively stable during this time.

We used the daily download numbers starting on Monday Feb. 28th (2011) and ending on Sunday March 10th. For each week we normalized the numbers to an average of 700 downloads a week (100 per day) and looked at the daily trends for these six working weeks. Once that data was complied, we removed two of the weeks with the greatest deviation from the norm from the analysis and came up with the following results as represented by these graphs:
<div><img id=”internal-source-marker_0.649012285284698″ alt=”” src=”https://lh6.googleusercontent.com/bHgix-XD-87wq64_dnEh4qnws3tqzGJjyIHS5LCXwRVZ2VDXbYZ6q5LZUVnL2WLRb6oTogVietOCVDgTNonUfxkWfWd0i8sHWRqsYnw5QTTUgh3T4io&#8221; width=”618px;” height=”273px;” /></div>
In tabular form this came out to the following matrix:
<table border=”0″><colgroup><col width=”*” /><col width=”*” /><col width=”95″ /><col width=”83″ /><col width=”*” /><col width=”*” /><col width=”*” /></colgroup>
<tbody>
<tr>
<td>Monday</td>
<td>Tuesday</td>
<td>Wednesday</td>
<td>Thursday</td>
<td>Friday</td>
<td>Saturday</td>
<td>Sunday</td>
</tr>
<tr>
<td>103</td>
<td>95</td>
<td>92</td>
<td>94</td>
<td>93</td>
<td>112</td>
<td>114</td>
</tr>
</tbody>
</table>
What’s surprising was how small the deviations for the entire data set were from this model:
<div><img id=”internal-source-marker_0.24265820160508156″ alt=”” src=”https://lh3.googleusercontent.com/-6WoTACyY_TljqUQ7BydYfUQkLI5K_7i–AM2XLCzpIUIghPUSpZ9qtvmfqdd79PsCU3GosatTwRZKYm0kJ2QxcZowminrZtTpJC3XcO_cWggK9MPgY&#8221; width=”621px;” height=”287px;” /></div>
The average deviation was approximately 4 downloads (normalized). Even with the ‘bad’ data folded back in, the deviations are still under 5 downloads on average.

As a check, we ran the same measurement for all six weeks to see if the model deviated greatly, even with the two previously removed weeks added back in. The values and deviation still indicated that the model was an accurate model of sales by day of the week. The averages were:

<img alt=”” src=”https://lh4.googleusercontent.com/xrCv_BpxJ7j7CJrR6g1RxDT09bykn6lC_wh2I2Hd16dfTU5HQ7sXpd9OpDa1DaMMFiqWzo3yVSF9vzDYhByTxxqtB8qLZHmsjKBAr3DDrgI3NdpCLEU&#8221; width=”625px;” height=”299px;” />
And the deviation (of the new averages) from the prior model (at the same scale) is:
<img alt=”” src=”https://lh6.googleusercontent.com/BbHD_gi972n0aZePsCUHf3_dtXhgviJxnKInzclSZjnQevRn2JZ4ADvxRSSjCrVh0fXhysEbFlkb5e-jYFh_EMfl2qTUx6TqY4u1T6QDL_T8xeJP-vs&#8221; width=”616px;” height=”285px;” />

What does it mean?

The implication is that you can scale your daily numbers by the inverse of these values to normalize sales number by day of the week. For example if you saw 6000 downloads on Friday and then 6500 downloads on Saturday, the normalized values would be 6451 and 5803 for Friday and Saturday respectively. Normalized Value being equal to the Actual Value divided by the model’s number for the day (as a %). The model for sales by day of the week would indicate that the sales for Saturday, given 6000 downloads on Friday, should be 7226 if all other factors remained the same.

So in the example given at the start, a change in a screen shot resulting in a 10% increase in downloads from Friday to Saturday (say from 6000 to 6600 downloads) … the corrected result is actually a 8.6% decrease in downloads when the day to day trends are used as a normalization. Better go back to the original screen-shot in that example.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s