Hacker News Reader

Comment by einhverfr | Hacker News Reader

I use a variety of spreadsheets but haven't touched Excel in years.

What a spreadsheet gives you, assuming you have enough math skill to operate one reasonably (i.e. at least a year of college algebra), is an ability to create tables of data following certain assumptions that project various scenarios. This is most helpful as a visualization tool. You could probably do something similar with R and some CSV files, but I find spreadsheets are actually really helpful here.

PS, I really don't like Excel, just because I have run into so many braindead issues with it. That doesn't say spreadsheets are not a very good tool for certain kinds of problems though.

View on HN

Of course, if you use python/R, you don't need to manually tweak a few assumptions in a spreadsheet and look at a few outcomes. You can instead just assign them a probability distribution and run a monte carlo simulation.

Doesn't that give me an average of the possibilities? I dont think people are usually looking for that... they usually want to specific cases, or graphs. both are very well supported in excel.

PS: I think excel probably has some support for monte carlo too. should look into it.

The output of a monte carlo simulation is a list of N (for N really large) of possible futures. You can average them, you can compute P(business makes a profit), you can do whatever you want with it.

If you want graphs you can spit out not just one, but multiple representative ones, credible intervals, etc.

Well, If I had a choice, I would definitely NOT want to assign a probability distribution and run a MC sim.

First because the probablity distribution itself would have to be pulled out of thin air, and make the results questionable.

Second because writing the assumption->outcome formulas is simply quicker than running an MC sim even if doing the sim is a one-liner (simply ensuring that the data is in appropriate format and testing once will take longer than needed).

Third (and main) because you don't particularly need the actual outcomes at all, the hands-on-tweaking of these assumptions and 'interactive learning' is the whole point of doing it all, you need to personally learn and feel the relation between these assumptions and financial results, and the 'result' numbers and graphs are just a side-effect and notes/docs to remind you later.

And if you need to convince someone else afterwards of your conclusions, then for 99% of audiences you anyway want to use 'specific plausible scenario story' (or comparisons of such scenarios) instead of a probability distribution coming out of a solid mathematical simulation of all possibilities; since it's well researched that the first kind of evidence works better in convincing homo sapiens about anything at all.

First because the probablity distribution itself would have to be pulled out of thin air, and make the results questionable.

Of course. It's just a more systematic version of manually tweaking parameters in a spreadsheet (which are also pulled out of thin air).

With libraries like PyMC, it's one extra step beyond writing an assumption->outcome formula. You write the same formula, but then set variables to probability distributions rather than floats .

And if you need to convince someone else afterwards of your conclusions, then for 99% of audiences you anyway want to use 'specific plausible scenario story'...since it's well researched that the first kind of evidence works better in convincing homo sapiens about anything at all.

Fair enough. This is solid dark arts, but useful.

If it can handle the same formulas, then I could imagine a market for a tool that takes an existing spreadsheet and does what you say - allows to set a probability distribution for any number, recalculates everything, and shows the distribution for resultcells on, say, mouseover.

Convenience matters in such things, and simply reading a 20x20 data table into any programming language already takes more time than just writing the formulas next to this table in excel.

I'm sure you could write such a plugin for excel with DataNitro or some such. The code is roughly:

    outputs = []
    for x in monte_carlo_input(param1_dist, param2_dist, ...):
        outputs.append( (x, f(x)) )

    histogram/whatever(outputs)

In contrast, the code to manually tweak formulas is simply:

    > f(10, 3, 0.07)
    12
    > f(12, 5, 0.09)
    -3.9

> First because the probablity distribution itself would have to be pulled out of thin air, and make the results questionable.

All aspects of financial forecasting in the initial stages are pulled out of thin air and are hence questionable. This is one of the big uncertainties one lives with when starting a business.

> Third (and main) because you don't particularly need the actual outcomes at all,

Bingo. That's the key. What you are looking for is not "will my business succeed" but rather "what do I have to do in order to succeed?" The tweaking of the model is the point, not the outcomes.

What for would you need "probability distribution and run a monte carlo simulation"?

Reply on news.ycombinator.com