I am looking for benchmark for forecasting accuracy within a financial organisation contact centre. Are there any industry standards etc.

Many Thanks

Stuart Wallis
Senior Business Planner
Halifax

Ah, my favorite topic. Stuart, assuming you referring to inbound call volume forecasting, here's some averages that we've encountered (~50 centers over the past 4 years):

Average Monthly Forecast Accuracy: +/-7-8%
Average Daily Accuracy: +/-11-15%

Best in Class Monthly Accuracy: +/-2-3%
Best in Class Daily Accuracy: +/-3-5%

Note that monthly figures are easier to hit, but accurately forecasting on a daily basis is really where the 'rubber meets the road' as far as operations are concerned.

No significant difference between the financial sector and other vertical markets that we've encountered, except clients in the financial vertical tend to have a greater focus on results.

Feel free to e-mail me if you have any questions, or if I can help in any way.

Brent

I can vouch for Brent's knowledge in this area, having received his help previously.

Although not in the financial services sector, I thought I'd post our own forecasting experiences.

Daily - average for the past 3 months is between +/5% and +/-7% (was nearly 10% before Brent's little bit of advice)

Weekly - +/- 4%
Monthly - +/- 2% to 3%

We had the joy of 4 different mailings of approx 120,000 (each go) ever month, at different points in each month and varying volumes, as well as no previous forecasts prior to August last year. So I'd hazard a guess that you might find the figures to be lower in more stable/cyclical environments and over time.

Hope that helps

Hello All,

Being new to this site, I have only found this thread.

I am the Forecasting Analyst for a utilities company. So, the comments about forecasting are very interesting.

Although we have a WFM tool, I am increasingly disillusioned with the 'regression analysis' form of forecasting it employs. If you throw a ball in the air, regression analysis would seem to suggest that the ball would continue upwards indefinately.

As such, I have contructed some sophisticated, and some not so sophisticated models for forecasting calls, using external factors such as bill volumes.

The frustration is that although much of the forecasting is very accurate, one day comes along every so often where the forecast is total out of whack, and for what seems no determinable reason whatsoever. Therefore, my aim is create a multi-factored tool where each factor is not independent of each other. However, this is a obviously a mammoth and complicated task.

My questions are; has anybody attempted this in the past who might be able to give me some advice from their experiences? And furthermore, is it possible to know what was Mr Preece's advice to Mr Tuck, that improved his forecasting accuracy two-fold?

Thanks.

Humm.......

Just a quick note.

You've just caught me as I'm about to leave for the weekend but I'll
try and put together some comments and idea's for you.

HTH

DvaeA

Hi, Tom - happy to provide a few tips, for what they're worth. Dave Appleby's response will give you some great ideas, as well. There's certainly more to forecasting than can be covered in a single posting, but we'll give it a shot.

It *is* a monumental task to account for various call volume drivers, but the good news is that it's been done before. I've got a half dozen clients in the utilities sector for whom we've created accurate forecasts, and while all markets are slightly different, the basic drivers are the same.

WFM tools are great at creating optimal schedules and helping to manage adherence. Forecasting accurately (+/-3% per day) is not usually their forte.

Regression analysis on a single line will do exactly what you said - assume a constant growth. However, a *multivariate* regression analysis works much better. I'll detail that in a moment.

In a basic sense, there are two types of call volume: base and variable. The base call volume is driven by the number of customers that you have. Variable call volume is driven by billing (you nailed that one, Tom), seasonality, holidays, marketing (particularly in a deregulated environment), etc.

The first line of your regression should be based on the total number of customers you have per month (going back as far as you can). The second line should be the number of new customers, the third should be the number of dropped customers. This will give you a base call volume.

The variable volume should include specific billing cycles for each day. Each drop is assigned a response rate (2-5%, depending on the season) and a response curve (percentage of total calls arriving week one, week two, etc). This variable volume is then layered over your base volume.

Seasonality is a big factor with most utilities. You can adjust for this by either changing the billing response rate on a monthly basis, or accounting for it as a percentage increase/decrease on the total volume each month. I recommend the former.

Holidays can stump a forecast if not accounted for. For instance, the Tuesday after a Monday holiday will be much busier (25-35%) than an ordinary Tuesday.

Also, there must be good communciation between your billing department and/or fulfillment group. A mistake on a batch of bills will generate unforseeable call volume, late billing drops will throw your forecast off by a number of days, etc. Other culprits for daily volume spikes are: higher volumes due to poor abandoned rate the previous day/week, newspaper articles or other (sometimes negative) press, surprise advertising (isn't it all?), power outages, etc.

There are some starters. Sounds like you already have a sound understanding of forecasting, Tom, hope this isn't too rudimentary. Please let me know if this leaves more questions than answers -

Brent

Hi,
Very interesting and informative topic that is really a help for many of the members in this forum, I guess! And Brent keep on posting, you really have a good knowledge and a wonderful way of explaining. Please correct me if I am wrong but I have doubts about Surprise advertising and Negative Press Articles will be the culprits for daily volume spikes, as I don’t think that they are really going to affect that a lot! And I think “changing the billing response rate on a monthly basis” is also a very good option. But you recommended “accounting for it as a percentage increase/decrease on the total volume each month”, I know that is also a good one but if I can have your valuable comparison on the same.
Thanks!

Tarun

Sorry Brent about the mistake of choices in Seasonality, as we both have the same opinion :).
Tarun

Ok,

A couple of days late but as promised some notes.

This to me is beginning to sound like a Viva Voce exam!

"Contact center Forecasting is an exact science. Discuss."

The answer to the question is both yes and no. (I'd make a good politician)

At the moment I'm running about a 95%-97% accuracy using handbuilt tools.
This however is only across 10,000 calls a week.

This is done with the emphisis on looking at the previous 6 weeks callflow,
looking at any market activity and re-forecasting as soon as the deviation slips
more than 5%. I also put into place an allowance for Indirect time (ie: time that cannot
be spent on the phone). This includes team meetings, breaks, training, appraisals,
project testing and coaching and buddying. The last one being important as we lose almost half
an FTE whilst they are buddying a new member of the phone staff.

The last 6 weeks data is analysed along with Average Handle times (AHT) by department then shoved
through a hand built Excel Erlang-C calculator. This takes into account the Indirect time calc as well
and returns a call level and planned staff level.

Ah I hear you cry but what about seasonality? It doesn't matter as the callflow is trend based on the
last few weeks we start to see peaks before they happen. Of course this means we can't forecast on
an hourly basis but with the allowance given and another trick we are acheiving <2% Aba and >90% in 16 sec
which is our corporate target.

When setting up shifts we use a model called middle office. This is a pool of Back office staff who are
Telephone trained and when I set up Rota's I use these as ad-hoc 1/2 hr cover. Where it's not worth putting
a full phone shift in but maybe 4 1/2 hr blocks need cover. It seems to work for us.

Regression analysis was mentioned so I'll explain this was the reason we started using the 6 week data and
old data just being used for in x month we got y% of our annual calls. Businesses change, processes change.
If you make one change that drops the repeat call rate it will affect everything. A 1% drop in the Aba rate will have a knock on effect on the Answer time and Call volume. If people don't hang up they don't
have to call back, call volume drops and performance improves. It's kind of a positive feedback spiral.

The next thing you need to do is keep an eye on what's actually happening. Look at how many staff you forecast
Vs how many were actually on the phone. Do this on a weekly basis for teams doing well and if required on a daily
if you've problems somewhere.

When it comes to billing cycles and suchlike you should be able to predict in advance the 'hit date' (Unless
you're dealing with the British Post Office) and the uplift should be remarkably static at each cycle.

Marketing is another problem as we use the UK Mailsort 3 (Delivery guaranteed sometime this year), this means
I have to look at the mailing size and type. Reserch the uptake and apply it to the mailing size and then guess
how it's going to hit. I was joking above but not by much. Mailsort 1 gives a guaranteed hit day, Mailsort 2
is within 3 working days of after a target day and Mailsort 3 is 7 working days. Unfortunatly the disparity of prices mean 3 is ALWAYS the cheapest option by a very long way.

My main method is to weight in a bell (Poisson) curve over the drop period and keep my fingers crossed.

I Hope this helps a bit.

If there's anything else drop me a line or ask here.

HTH

DaveA

I have built a few Forecasting Models that I use in combination with one another - I find it's this combination of methods that helps refine the forecasts to the accuracy's I require.

Like Dave I have a short term Forecasting Model based on the last six weeks actual volumes, this is a fairly robust model and as Dave said constantly updates with the latest trends, however there is no trend or seasonality that can be used to accurately extrapolate this model further than one period (or Week). This leads me on to my next model...

I have a medium term Forecasting Model that I use to forecast on a 3 month rolling basis based on the past 12 months data. It uses Holt-Winters triple exponential smoothing method, which I optimise with a genetic algorithm coded in VB. This model introduces trend,slope, & seasonality, I find the accuracy I require can be maintained for upto 3 month periods. To forecast on a more long term basis I have another model...

My final model is an Explanatory model which I have built to look at the long term. This model looks at all the call drivers for the different customer types we have, for each type I can assume propensity's to call which I then apply to a forecasted customer base.

As mentioned I have links between all three of my models, the assumption of continuity cannot can not always be applied, sometimes Marketing will schedule a new campaign at a time of year which was previously quiet, billing cycles might change, etc, so any related information I have that would affect the forecast, I cascade down from my explanatory model to the time series models.

Even with all my models set up, they still require a lot of manual interaction, a forecast is only as good as the data it uses, I have an integrated noise reduction algorithm for the time series model, however I still manually check over the data. The Explanatory model requires not only heavy analysis, but also strict process and communications with other departments.

I guess the next step on from what I have, would be something like IMAJ's Neural Net Forecast Model, by all accounts a complex but neat piece of kit!

sfk at softfurrykitty.com

Need your assistance in forecasting. Hope to hear from you.

Thanks,
SFK

just an add in to the excellent advice already given - Brent mentions factors such as post bank-holidays, marketing campaings etc, but there is another levelk of detail to add here if you want to get your 15 minute service intervals more accurate: television.
On the large scale, this involves national sporting events (look at your July 02 volumes), major charity events such as Children in Need, etc. On a more detailled scale, Eastenders, the channel 4 news, and other programmes with similar viewing figure will also have a direct impact especailly when a "big story" episode is billed - the specific programmes that will influence your volumes will depend on your demographic. You can see the effect of this by comparing call volumes in the 15 minutes during a programme compared to the inbound volume just after, where you will often get a surge.
When you use historical data for forecasting be sure to factor the mitigating events back out if they are not reoccuring e.g. if you rellied on last July's inbound volumes for this years forecast I bet you had a hard summer! For this reason it is important that volumetrics are accompanied by relevant notation, or that you produce adjusted volumes to account for special events.
does everyone else expect 15 minute service interval management (including forecasts) or am i too demanding?