Monday, May 9, 2011

Reference Class Forecasting

Bent Flyvbjerg has a long track record getting at the root causes of cost and schedule estimating errors in large scale projects, particularly those in the construction and transportation domains. His work is both theoretical and pragmatic, reflecting his former position as professor of planning in the Department of Development and Planning at Aalborg University [Denmark] and his current position as Research Director and professor of major program management at Oxford University [UK]

One of his favorite targets is the Sydney Opera House, notoriously difficult to build, some 1400% over budget, but a priceless source of  civic pride; and considered by most to be an architectural masterpiece. [Business value vs project value?!]

I ran across some of Bent's papers in a search I was doing on estimating.  One, "Design by Deception"  is a litany of major failures with some insights to their problems, mostly cases of looking the other way and choosing to believe the unbelievable. It's definitely worth a scan.

However, my attention was drawn to a paper he wrote the the PMI Project Management Journal, somewhat strangely entitled "From Nobel Prize to Project Management: getting Risks Right"

In spite of the title, the theme of the article is a practice named "Reference Class Forecasting".  From one view, RCF is just cost history applied to parametric model-based estimating, a method that's been around forever.   However, Bent and his co-authors spin it a little differently.  Their idea is given in 4 steps:

Step 1: Form the 'reference class', a collection of similar-to projects for which there is both history and reasonable insight to the history so that adjustments for present time can be made.  [Bent never did say what the simliar-to projects might have been for the Opera House]

Step 2: Develop a true distribution of the reference class, and from that distribution calculate the cummulative probability.  [Actually, they may have done it the other way around, but the main point is to come up with the cummulative probability].  They call the probability curve, developed from reference class, "the outside view".

Step 3: Develop the "inside view".  The inside view is a traditional estimate by the project team.

Step 4:  Adjust the inside view based on the probability of historical outcome from the outside view.  That is, develop a forecast using the reference class probability confidence curve.  In effect, according to policy or doctrine, or other direction, pick a confidence limit, and then adjust the inside view to have a corresponding confidence.

Obviously, the objective of RCF is to improve the confidence in the final estimate.  Along the way there are a couple other objectives addressed.  One is to overcome "delusion" brought on by "optimism bias", a phenomenon studied by Tversky and Khaneman.  A general statement of such a bias is that indiviuals with optimistic outlooks tend to underestimate risks; the corollary holds: depressed individuals tend to overestimate risk effects.

The other is overcome--or least provide the ammunition,--"deception" brought on by political neccesity.  Of course, in the latter case, legitimate accuracy is not actually politicially convenient.  Many times, the better data is 'buried'.  Shocking!

The good news is that delusion and deception tend to be counter-acting.  When one is in accendance, the other tends to retreat.  Obviously, as an academic, Flyvbjerg has some appreciation of the politics of delusion, but he campaigns against it nevertheless.  On the other hand, delusion seems to be the easier target to shoot down.

Bookmark this on Delicious