How to teach the Forecasting Algorithm?

Each day is similar to a previous day. The question is to what extent and to what type of day it is similar to. What weight should you assign to historical data for the best prediction?

An average Wednesday is similar to an earlier Wednesday, but it also has to be taken into account which day of the month it is, because there is a periodicity by month in inbound traffic.

That is why we use seasonality values for better prediction.

Each day has 3 different rates of seasonality: weekly, monthly & annual

Seasonality settings for the forecasted period

Each day has three attributes according to weekly daily and annual seasonality

e.g. Seasonality of 25th December: 100 % annual, 0 % monthly, 0 % weekly

For the next period we give default values of seasonality for each day that can be changed by the supervisor.

Erlang calculations for certain skills

The Erlang parameters can be different by skill and by days of the week.

With the given SLA values the Forecasting module calculates the required number of agents by skill.


The result of Forecast

The result can be seen on this graph. The predicted values can be compared with the facts (green curve) if the facts have been already uploaded.

The date of this snapshot is 9th of February because we see the factual data till 8th of February. (The data of the previous day are received during the night.)


Altering effects

In certain cases we know in advance that i.e. we will launch a TV campaign that increases the incoming traffic by 20%. So we can increase the predicted traffic by this value.