1
$\begingroup$

It's a bit untypical question I guess, but I hope you can help me. I know a little bit of statistics. I'm not a specialist, but I find it really interesting, so I learn in my free time. I need your opinion on the following conflict.

I need to prepare sales forecasts for an e-commerce website. Imagine it as Amazon (it's not Amazon).

  1. My approach:

    • I took past sales data. First I analysed seasonality and trend.

    • I also analysed the main factors influencing the sales

    • I tried out several forecasting methods (ARIMA, Holt-Winters, neural networks), but in the end a regression model based on the factors identified as main determinants was the most convincing one. (I didn't have enough data for neural networks). The forecast resulted quite good.

  2. My coworker's approach:

    • We should take the data on "sessions" (i.e. how many times the website is accessed by customers daily), let's say it's 2k
    • We should then take the data on the "conversion rate" (how many sessions on average result in a sale, e.g. 20%)
    • This will allow us to get our sales forecasts: by multiplying the number of sessions by the conversion rate. In this example, it's 2k*20%=400.

I find his idea problematic for several reasons:

  • First, we can only get data on the conversion rate and sessions for NOW. Not for the future. So if we need this data for future, we need to forecast it, adding complexity to the model
  • Data on sessions and the conversion rate is strongly volatile - it's just volatile as the sales data - adding complexity to the model. Basically, we would have 2 volatile variables instead of one.
  • The determinants of the conversion rate and sessions are not necessarily the same as the determinants of sales

Is there any possibility that his way of approaching it - i.e. making forecasts on the basis of the data on sessions and the conversion rate - is better than mine?

I want to be 100% sure my approach is correct. I'm keeping an open mind and if his idea is better, I want to know that.

$\endgroup$
0
$\begingroup$

One possible way where your colleague's approach might perform better would be if the two "components" had very different drivers. For instance, I suspect that there are strong for sessions, with people accessing the site differently over the day and between workdays and weekends. However, there might be much less seasonality for the conversion rate. This might be driven much more by things like whether a given session results in a basket that is large enough for, say, shipping costs to be waived. So, yes, it might indeed be possible to forecast both separate components very well, so that the final product is a better forecast than a "direct" approach.

As a matter of fact, when two different approaches to forecasting are in play, it is often best to combine the two, e.g., by calculating the average of the two forecasts. (With the added complexity that you now need to maintain both algorithms. This may or may not be worth an increase in accuracy.)

You could compare the forecast accuracy of your and your colleague's approach (or the average) by using a holdout sample.

$\endgroup$
  • $\begingroup$ There seems to be no seasonality for the conversion rate. The conversion rate is much more changeable than the sessions and sales. When you look at the sessions and sales, you do see some patterns (not the same patterns of course). When you look at the conversion rate it seems much more random. Another thing: we need forecasts several months in advance, so forecasting the conv. rate on the basis of what people have in their baskets wouldn't work - I don't think many people keep products in their baskets for months. $\endgroup$ – BigMadAndy Aug 24 '18 at 14:31
0
$\begingroup$

Usually sales data and other monetary data has more fluctuation than number of unit data, because the value of money is very volatile. So, forecasting the session should have less variance (should because there are many exceptions), and most of the seasonality will go there.

The "conversion factor" (the propensity to buy) is however a tricky beast. Try to correlate it with the your "main factors influencing the sales" as well as the size of the buy, and the age of the basket.

Now, to settle the competition with your colleague: if the conversion factor is assumed constant, the two approach have the same quality. Not the same results of course, but you will have no way to tell which one is better.

$\endgroup$
  • $\begingroup$ That's the very point, however, isn't it? As I write in my post, the conversion rate is strongly volatile. It's much more volatile than the sales and sessions, ranging from 20% to 80% for one product even within a day. Obviously, if the conv. rate was stable, both ways of counting it would be good. But it's not. I also have some methodological problems with this approach. If sb asked me to forecast the sales of beer in the supermarket, I wouldn't go and observe how many people are watching shelves with beer and calculate the conv. rate either. $\endgroup$ – BigMadAndy Aug 29 '18 at 20:53
  • $\begingroup$ Yes, that's the very point. Unless you can find an explanatory variable to reduce the "conversion" rate (which will bring you on a long journey in customer behavior), and the direct method gives good results: use it. $\endgroup$ – AlainD Aug 31 '18 at 8:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.