Capturing expert inputs about price variation in the future I am creating a profitability model for a proposed chemical manufacturing project and some of the parameters that go into this model are subject to uncertainty. 
e.g. The price of Styrene (a feedstock), or the cost of electricity, cost of coal etc.
The goal is to not only estimate the profitability of the project but include some information about the variability of the estimate. Perhaps 95% uncertainty intervals. 
One way to capture this is to actually ask the relevant experts. e.g. The key coal suppliers. Or people from the electricity distribution business.
I know this is not foolproof but it's a start to quantifying the uncertainty. 
My question: What's the best way to capture information from these "statistical laymen" so that it can be translated into some sort of useful measure of variation of the underlying price? 
e.g. 
I could ask:


*

*What do you foresee as the most likely price of Styrene one year from now? 

*What do you foresee as the standard distribution of Styrene price variation? (but that does not make much sense to most people)

*What do you feel the lowest or highest price of styrene could be? 


etc.
My hope is that this info can be somehow mapped to a distribution of likely values of that parameter & then drawing from these distributions (hopefully uncorrelated) I can generate some sort of distribution of the final profitability. 
Even if I assume a normal distribution for all parameters, what are good framings of questions to ask a non-statistician whose answers can then be mapped to a mean & standard deviation of the distribution to use?
Any thoughts how to go about this?
 A: There's no right answer to this and you will have to be careful about how you synthesise these opinions with any hard data that you subsequently record. With those caveats out there, I'd recommend asking questions about the probability of prices exceeding a certain amount. This can be translated into subjective prior quantile estimates for modelling purposes.
In other words I might ask:
"What's the probability that next year's styrene prices will be X% higher than this year's prices?"
A: There is now an extensive literature base and empirical results showing that particular questions asked in a particular order provide the best results in terms of expert judgments that reduce overconfidence and systematic bias. This short and accessible book and this longer and more comprehensive text are great starting points for getting into this field.
While @RichardRedding is correct that there's no "right" way, there are certainly bad ways. In brief, the best practice for this sort of elicitation exercise is a four-part question. After thorough framing of the question (i.e., being clear with your experts about what values you'll be asking for, the format of the questions, and laying out all of the base assumptions), you ask questions in this order:


*

*Think of all of the reasons the value might be low. Realistically, what is the lowest plausible value for price of styrene in on January 1, 2019? [Be sure to document the reasons the expert is considering.]

*Think of all of the reasons the value might be high. Realistically, what is the highest plausible value for price of styrene in on January 1, 2019? [Again, documenting these reasons.]

*Realistically, what is the most likely value for price of styrene in on January 1, 2019?

*What percent of plausible values do you think are captured in the range you provided [ideally, as a percentage over 50%].
Once you have these estimates, you can parameterize a variety of distribution families depending on your needs. Part 4 provides a credible interval that is defined by parts 1 and 2, and allows you to extrapolate to the appropriate distribution parameters. Part 3 provides the modal value, which, depending on your distribution family, may or may not be the mean. In my experience, experts rarely provide a symmetrical distribution, and so normals usually fit their estimates poorly. Triangular distributions are simpler, and in my opinion don't lose much information unless you're concerned heavily with the tails (i.e., very low probability events). Based on this information, you can provide feedback to your experts about the probability distribution they've defined. In this figure, the range provided by the expert is in blue, and the full implied distribution is the black line:

A: You asked " What's the best way to capture information from these "statistical laymen" so that it can be translated into some sort of useful measure of variation of the underlying price?" 
In the spirit of your question , we have recently developed procedures to introduce "Delphi-Type" forecasts as inputs to Transfer Function ( REGRESSION ON STEROIDS !) forecasts https://autobox.com/cms/images/dllupdate/AutoboxUsersguide.pdf slide 94 ..
In this way the history of predictors can be used to form the model and the appropriate response function to both known predictors (inputs) and deterministic latent predictors AND then forecasts can then be made using the equation and monte-carlo like forecasts eminating from a user specified probability range for one or more delphi-type inputs all in an optionally automatic way.
