Can you develop an econometrics model for stress test purpose only focusing on 2008-2009 data? I have become aware that a group at a large corporation is developing an econometrics model to forecast sales of their product.  They are using this model solely to estimate sales in specified stress test economic scenarios where they are given what the economic environment will be like, including real GDP contraction, rising unemployment rate, etc... out to 2016.  Because of the nature of those scenarios, they think the most proper way to construct this model is to focus solely on the 2008-2009 period capturing the main period of the recent financial crisis.  They have monthly data, so that gives them 24 monthly data points.  Given that GDP's frequency is really quarterly, on this one variable it gives them only 8 true datapoints.  But, they extrapolate it into 24 month observations. 
For the record, if they chose to, they have good internal data going back to 2001 and up to the current period.  But, as mentioned they decided to focus instead solely on the 2008-2009 period.  
I will also answer this question as I have built many such econometrics models.  And, I invite others to debate and rebutt my answer... and to post your own better answer. 
 A: The points you are making are valid, but there are also arguments that if not counter to yours, they create a dilemma:  
When trying to estimate and forecast extreme cases, incorporating information from "normal times" may "average" your predictor, which would then be more reliable to estimate long-term trends rather than short-term (and severe) fluctuations. Models that describe well both these aspects are still not available, because we do not yet understand well how "normal times" breed their own crises (in "normal times" I include the concept of a business cycle - a crisis is something much more severe).
One could build three models: one based just on "crisis data", one based on only "normal times" data, and one based on both data. Comparing the three in terms of their forecasts would be very valuable. Also one could  implement "model-averaging" on the two "pure" models and compare its forecasts with the "both kind of data" model forecasts.
Since they are a large corporation and only in it for the money, this multiplication of resources allocation to estimate their sales can be justified -and financed.
A: An interesting and topical issue in risk modelling. In my experience, risk models for credit which are developed on shorter periods of data tend to produce unstable coefficients - cross validation and/or out-of-time testing for model performance have usually shown this to be the case.
For stress testing of credit risk models, we are concerned with estimating probability of default (PD, probability of customer not repaying a loan) and loss given default (LGD, proportion of loan lost in event of default) during downturn macroeconomic conditions.
In terms of regulatory perspectives for stress testing, the Basel Committee on Banking Supervision (BCBS), which can be considered the central bank of central banks, indicates a minimum of 5 to 7 years of data (dependent on model and portfolio type) for model development, unless strong evidence is shown that more recent data is more predictive. Banking regulators typically adhere to these time periods as part of Basel II and III standards for calculating regulatory capital.
Additionally, the time period of 2008 to 2009 would seem to be somewhat short as the financial crisis has persisted for a longer period in some countries, e.g. UK where macro economic conditions worsened in 2007. It could be argued that the crisis is still ongoing in some parts of the world.
A: When building such econometrics model, you want to grab as much historical data as you can.  That's so you can observe the relationship between your dependent variable and the macroeconomic ones for a long enough period that you have enough data points so that such relationships are somewhat reliable.  The observed correlations and regression coefficients will both be more statistically significant and meaningful than otherwise. 
You want to observe those relationships over full economic cycles not just during a crisis.  Every crisis or downturn is somewhat different.  And, that is especially true of the 2008-2009 financial crisis.  If you focus solely on this period, the variables relationships were probably unstable and not so representative of any other economic environment except for one identical to 2008-2009 which may not happen for another 80 years or so (the most recent parallel is really the Great Depression that started in 1929). 
When looking at stress test scenarios.  They go out several years (2014, 2015, 2016, etc...).  As structured, the first year or two they do reflect a severe and protracted downturn.  But, sometimes by the second half of the second year and onword... they actually reflect a recovery.  The latter can be structured as a slow one or sometimes even a really strong one.  If within your model you have only the 2008-2009 as a learning history, the model will have no idea on how to estimate sales during a recovery or a more normal economic climate.  In other words, your standard economic climate could become "out of sample" since you have never captured such a climate in your 2008-2009 learning history.  For an econometrics model... this spells disaster.  In this specific situation, I would definitely grab the entire data set from 2001 to the present.  As is, this data set is actually quite short to develop a reliable econometrics model. 
A: I think a more precise goal of the analysis needs to be articulated. If the objective is to come up with revenue projections in worst case scenario only, then I think it is appropriate to only use data from recession periods. 
If you are mixing in data from other non-recession years, you are liable to get more optimistic results and not answer your original question of what revenues are going to be like in recessions. 
Ideally of course you want as much data as possible to draw inference, but it's not proper to add irrelevant data just because it's a hassle to use small sample statistical methods.
I think it is inappropriate to consider 7 years of data like the Basel Committee recommends For one thing, these people are not banks. Second of all, reserve requirements and people's demands for loans likely aren't comparable with people's demand for whatever product the questioner's companies are selling. 
A: 24 points (2 years) is too small data to get a reliable model in my experience with this exercise.
But even if you would have a 5 years of stress data - your probably still don't want to use only stress period to fit a model. Because your scenarios' pack contain both stressed and normal scenarios and even stressed one have some recovery part in the end.
So I would fit a model using all historical data - 2007/2008-crisis and normal periods; in this way the model will fit both stressed/normal dynamics and then your future scenarios - stress and baseline - is what will drive forecasts and what differentiate stress/normal forecasts rather than historical fitting data.
