In Part 1 of this series, we dis­cussed the appli­ca­tion of fore­cast­ing mod­els to esti­mate miss­ing data points. In this part, we will dis­cuss iden­ti­fy­ing future trends and anomalies.

Fore­cast­ing Appli­ca­tion: Iden­tify sig­nif­i­cant traf­fic patterns

Almost all trended graphs include peaks and val­leys, which cap­ture our atten­tion. These anom­alies stand out to us and raise the ques­tions “What hap­pened?” and “What should we do next?” Before div­ing into the details, pause and deter­mine if the change from the trend is expected, or if it is some­thing of sig­nif­i­cance.  If the change in the trend is unex­pected, we can employ sta­tis­ti­cal mod­el­ing to deter­mine if the change is sig­nif­i­cant, mean­ing do we need to raise an alarm and do some­thing about it. Time-series fore­cast­ing mod­els will not only allow us to view expected out­comes, but the mod­els pro­vide an upper and lower con­fi­dence bound. If a data point lies out­side one of those bounds, then that is a good data point to start analyzing.


While prepar­ing for the end of year hol­i­day sales, an online retailer wanted to get an esti­mate of traf­fic and sales vol­umes to their var­i­ous prod­uct pages. With the data avail­able in Adobe Ana­lyt­ics, we used a time-series fore­cast­ing model (actu­ally two to val­i­date, an ARIMA and a Holt-Winters decom­po­si­tion) to fore­cast expected traf­fic through the hol­i­day sea­son. To everyone’s sur­prise, one of the prod­uct lines was esti­mated to have a sig­nif­i­cant drop in traf­fic. Other prod­uct lines had either increas­ing or decreas­ing trends, but with lim­ited resources, the team could not devote time to each one. Through the time series model, the prod­uct line with the sig­nif­i­cant decrease was iden­ti­fied and action was tak­ing to improve traf­fic and con­ver­sion into that area. Here is what we did:

  1. Iden­ti­fied rel­e­vant traf­fic sources and vari­ables for modeling.
  2. Iden­ti­fied a good time-series model for each source (part of this series iden­ti­fies poten­tial fore­cast­ing methodologies).
  3. Added the mod­els to Excel dashboards.
  4. Iden­ti­fied the data points below the trend lines and con­fi­dence boundaries.
  5. Iden­ti­fied the data points and vari­ables highly cor­re­lated with this prod­uct line – This can be done through a sim­ple cor­re­la­tion in Excel, but be aware that just because two items are cor­re­lated, does not mean that one event caused the other.
  6. Took action on the cor­re­lated activities

To help with the action step, we only looked at areas that we could influ­ence such as paid mar­ket­ing chan­nels. If we saw a rela­tion­ship with social refer­ring traf­fic, that’s great, but may be more dif­fi­cult to act on as opposed to email or paid search.

Fore­cast­ing is a great tool to add to your ana­lyt­ics tool belt, and it is eas­ier to apply than you may think. Remem­ber to use a good rep­re­sen­ta­tion of your pop­u­la­tion as your sam­ple data for fore­cast­ing. Also try to have a min­i­mum of 18 months of data – sta­tis­ti­cal mod­els tend to work bet­ter with more data, and some­times the math will not work if you have less than 12 months.

As a final thought, remem­ber the con­text. Out­side vari­ables may be influ­enc­ing your data. The idea is to not let mod­el­ing and sta­tis­tics replace your cur­rent thought process and intu­ition, but sup­ple­ment it to strengthen your analysis.

For addi­tional infor­ma­tion, or to speak to a mem­ber of the Pre­dic­tive Mar­ket­ing team about other ben­e­fits of data min­ing and applied sta­tis­tics in Dig­i­tal Mar­ket­ing, con­tact your Adobe representative.