This is an edited version of an e-mail conversation between a new user of Forecaster and the author, in March 2009. The dialog covers basic information on how to use the features of the Forecaster program, with an emphasis on the Withdrawal Decision Rules simulation.
Q. Thanks for your retirement program. Your program is the best retirement planning tool I have ever seen. I am approaching retirement next year and have been perusing just about everything on the subject. The market has been so bad that my starting capital has been reduced by about half from a year ago. So now I'm trying to figure out the best way to approach the future returns and inflation estimates. Based on past history the market has made big recoveries after recessions, but I don't know how to input that - or even if it will do so again this time. The retirement forecasting procedure of your program seems to be to first run the analysis, then run the Monte Carlo, and then tweak it with the withdrawal decision rules simulation. And then add innumerable permutations of those variables. What could possibly be better? Thanks so much for sharing your work to make retirement so much clearer for anyone with enough luck to find your program. I found it by playing around in FIRECalc . Also your videos are terrific. I learned a lot of ideas by watching the three videos over and over and finding your ideas and others that popped up during the viewing. Now the only issue I have is how to leave the program alone long enough to get other stuff done.
A. Thanks so much for your testimonial. I really appreciate hearing from someone that likes my program and uses it. I'm especially glad the videos worked for you. They are a new device for me, because I figured that the withdrawal decision rules documentation would be too complicated, and a video would be the best way to tutor users. Now I think that I might add some more videos to explain some of the other esoteric features of the program.
Q. I agree that the videos are the best possible way to share your program's seemingly unlimited features. On a very personal level, but also as a current view of the economic times we are in, do you have any ideas on how to model retirement now with beginning values "temporarily" diminished at the beginning of retirement (a little different than the typical view of diminished returns early in retirement versus later). Maybe show a return of, say 35%, in a given year to show an average recovery based on prior recessions. But I wonder how unrealistic that would be if such a typical recovery does not occur this time. I would appreciate your thoughts on this economic meltdown and how best to model retirement going forward based on this absolute fact (unfortunately, a real rather than modeled "perfect storm"). I think this would be a valid basis for a video on this very real situation.
A. The Demo3 video on this Web page shows how to model retirement with beginning ROI values "temporarily" diminished at the start. The demo shows setting the ROI to 0% for the first few years, then setting ROI back to 9.2% for the remaining years. You could set any year to have a ROI of 35% and then the following years back to your normal ROI.
I can tell you how to use the program, but I'm not a financial or economic person, so I offer no opinion or advice in those areas. I can recommend this blog: http://www.mebanefaber.com/ as a source of information.
Q. Thanks for your info about the Demo3 video for modeling the current economic situation. Sorry to bother you again, but is there a way to show the yearly payout amounts computed by the Withdrawal Rules Simulation? Your program shows a graph of these payouts, but I have not been able to see the actual payouts by amount per year, like your program does show for the Plan Overall Results printout.
A. The short answer is no. That is because the lines you see on the graph are based on the probabilities from running 100 (or whatever value you specified) simulations using different ROI and COLA for each. So there are 100 (or more) different yearly payout results. There is no single series of yearly payout amounts that is meaningful enough to commit to a table. That would imply too much validity to those numbers. The whole thing is just an exercise to show that by using yearly adjustments you can conserve your capital. The parameter values that you determined show the starting withdrawal value that has the best probability of producing the desired plot of yearly payouts over the future. My advice is to run the whole exercise each year, using up-to-date numbers, then just use the first year's number as your guide for that year. Repeat each year.
Q. Thanks for your reply. I understand the yearly expense funding graph only shows probabilities and not meaningful numbers.
So, here's how it works, as I understand it:
1) The Calculate function shows whether there are any problems with the basic inputs,
2) The Monte Carlo function shows the likelihood of the model working out with the extra standard deviation value inputs (in most cases, the results are high likelihood of failure),
3) The Withdrawal Rules Simulation can be tweaked until the results are successful again to work-around the negative results from the previous step.
At least, that's how it has worked out in my situation.
In the last part of your answer, do I understand that I should use the first year's payout number in every year thereafter, but change the other variables to up to date numbers? For example, if I retire at age 65 and run the program successfully then, when I turn, say 70, do I still use the age 65 first year's payout at my age 70, even though by then the actual payout would probably have changed. Specifically, assume first year's payout of $80K at age 65, which the program shows to be $93K at age 70 with inflation alone, do I use the $80k at age 70 and calculate forward (or do I use $93K, if that is the payout then)?
A. Yes, you're right. The calculation function checks that the specified plan has no errors. It allows you to confirm that the accounts' accounting and the flow of funds are how you intended. (Most useful for estate planning of gifts, insurance premium payments to trusts, beneficiary payouts, etc.) But the use of average values for the calculation variables gives an optimistic forecast.
The Monte Carlo simulation uses a random selection from a normal distribution (a bell curve) of ROI and COLA values derived from your specification of the standard deviations of those values. This gives a more realistic forecast. But the variances in ROI and COLA values may lead to a lower probability of the desired outcome (not going bust).
Now, about the Withdrawal Decision Rules (WDR) simulation. To be truthful, I don't have a good idea how to use the WDR. If you read Klinger's paper , "Using Decision Rules to Create Retirement Withdrawal Profiles" , you'll see that he describes a simple case where the portfolio has no minimum required IRA withdrawals, there is no addition income such as a pension, and taxes are not considered. He also has created a program that can be found here: http://www.b-k-ind.com/
When I read Klinger's paper , I thought that it would be an interesting programming exercise to add his WDR ideas to my Forecaster program. I wanted to see if his ideas would work in a real world environment that included all the other constraints that would affect the actual net amount of money for retirement living expenses. So that is why my implementation is what it is. I can't say that I'm happy with it because I think it's hard to use and the results are hard to follow.
Klinger focuses on the amount withdrawn from the portfolio. My implementation focuses on the net amount needed to fund retirement living expenses. The portfolio withdrawal is only enough to make up the difference after the other funds from pensions and minimum required IRA distributions. Everything is based on after tax amounts.
I was surprised that I had to modify Klinger's logic to allow the DWR's Fall value to be "blank" (so that the Raise is always applied) to get the Yearly Expense Funding chart to show a flat line for the median value. Besides not being a financial or economic person, I'm not a mathematician either. So I don't have the skills or tools to analyze what is really happening, but I'm guessing that because of inflation the relative contribution of the fixed pension income is reduced over time so that the contribution from withdrawals has to be constantly increased to make up for it.
And as I pointed out to you, there is no resulting set of year-to-year instructions to guide you in the withdrawals required to meet your goal.
But here's what I think the WDR process can be used for. After you have supplied your estimate of the ROI and COLA standard deviations, and the mean values of ROI and COLA you expect to see, then you adjust your retirement living expense and the WDR parameters to achieve a flat chart line for your yearly expense funding with only a 1% chance of failure. Now, based on that simulation result, which gives an acceptable probability for your future forecast, you have a number for the current year's total portfolio withdrawal. This is what you withdraw to add to your other sources of income for the year. Then, after paying taxes, you have your funds for this year's living expenses.
But you can't really expect the data that you entered this year to be the same for the future years. So the thing to do is repeat the whole exercise next year. Change everything to reflect current conditions: your age, your account values, tax rates, the calculation start year and your retirement start year (both the same if you're retired). At that time use the best available information to make new estimates for ROI and COLA standard deviations and the mean values of ROI and COLA. Adjust the WDR simulation parameters to achieve the flat chart line and the 1% failure rate. Now you will have a new number for the current year's total portfolio withdrawal.
By doing this each year you are using the WDR logic along with the best current information to calculate future probabilities in order to determine the optimum portfolio withdrawal for each year. I think this is the best you can do.
Q. It took me awhile to understand all the info in your e-mail, so haven't replied before. Thanks a lot for providing such a complete description of the process and functionality - I will use it as a guide as I work with your program. It's discouraging to get an error-free result in the Calculate function, and then to run Monte Carlo and see it all wiped out. I really like the WDR results, because I can tweak the inputs enough to achieve a 99% success, but I just hope the Monte Carlo isn't the more accurate result.
Why is Monte Carlo so darn negative? For example, the Calculate function results in $2,283,992 after 40 years, but Monte Carlo shows only 81% success, and WDR using the default inputs from the Klinger article (all rules checked and fall/raise and exceeds/cuts at 20/10) achieves 100%. Which result is the best for retirement reality? I noticed in the Calculate printout that the parameters show the lower of the two Standard Deviation numbers. Does Monte Carlo use the higher of the two numbers and that's why the results are so negative?
A. In regard to your Monte Carlo question, the two Standard Deviation numbers mean the same variance. One is one standard deviation and the other is two standard deviations.
I think you can understand the simulation process better by reading these two Web pages:
In your case, your standard deviation values are high enough so that in many of the simulation runs, the sequence of random numbers causes you go broke before the end of the plan.
You can get different Monte Carlo results by changing the random number seed in The Monte Carlo Simulation Parameters data entry screen. This just causes the random number generator to create a different sequence of numbers, which is analogous to a different set of economic and investment circumstances.
Try the seed number 70765 for instance. In my own plan that gave me a Monte Carlo result of 100% chance of success, whereas the original seed number of 700 gave me a 24% chance of going broke. But this just shows that the forecast is always just a question of probabilities, and a random number sequence will give different results depending on what those random numbers are. Neither result is "better" than the other.
Q. The Monte Carlo links you mentioned are excellent resources and clarified this concept greatly. In my Monte Carlo runs, I continue to achieve about 80% success, no matter what random number seed I input, even with your 70765. I am modeling my ROI and COLA and Standard Deviations after the Klinger article on 08/2007: ROI 9.62%, SD 19.5%/9.75% and COLA 3.1%, SD 4.16%/2.08%. I am using the 9.62% since my portfolio is nearly all equities, and that was the Klinger S&P500 average for the period 1926 to 2004 (AXA). Could these Klinger numbers be causing the problem why I cannot get better than 80% success? Is there a better way to obtain long-term averages for ROI and COLA to input for the Global Plan Parameters?
A. The numbers from Klinger : ROI 9.62%, SD 9.75% and COLA 3.1%, SD 2.08% are as good as any for your long term assumptions.
For each year of your retirement, the Monte Carlo simulation will randomly select a ROI value from a Gaussian distribution with a mean of 9.62% and a SD of 9.75%. This means that for each year there is a 16% chance that the ROI will be less than 0.17%. A bad run of luck (the random numbers) can wipe you out.
In all the cases that you have tried, apparently the numbers have simulated a sequence of years with low ROI so that you run out of money. Changing the seed should initiate a different sequence, but the new sequence could still have a string of bad luck at the wrong time
Are you still using only 100 trials? Try running 2000 (or more) trials. Then maybe the random numbers will turn in your favor over the longer period. But I still think trying different seed numbers is the best way to get different results of the probabilities.
Q. I finally got a Monte Carlo 100% success by using seed number 99 with 1000 trials - But I ran the exact same setting again and got 88% success. Did the ROI and COLA numbers get scrambled in my favor on the first run and re-scrambled into a bad sequence for the second run? When do you think a person can say that's as far as I can go and be content with the outcome?
Sorry if this question is rudimentary, but how do you contrast random numbers like your program with historical numbers like the FIRECalc program? Is it like asking to contrast basing your future on history or rolls of the dice?
A. I can't explain why you got 2 different results with 2 Monte Carlo simulations both using the same numbers. Using the same seed is supposed to generate the same sequence of numbers each time. The random number generator is provided by Microsoft, and that is the way it should work. That's how it works for me.
As I understand it, FIRECalc uses historical yearly values for ROI and COLA. It then starts at some point in the past and runs through the historical data to get the numbers for your retirement span of years. Then it starts again, displacing the start by one year. So you get 100 simulation runs using a different sequence each time. But each sequence is an actual sequence from history. If you start just before a bear market, you'll have a bad result. If not, your result will be better.
Some simulations I know about even do it another way. Instead of keeping the yearly numbers in year order, the system will still use real historical numbers, but will randomize the yearly sequence and then use that series of random yearly values to calculate the simulated result. Still real numbers, but not serially related.
Using a random selection of ROI and COLA from a Gaussian distribution tries to simulate the variances seen in these historical values. It is a purely mathematical solution to the exercise of predicting possible yearly values.
So, is the choice to be using history or rolls of dice? I think relying on history is itself a roll of the dice. Is there any historical record of an economy like the one now where all supposedly uncorrelated assets fall together across the board?
I think your results are giving you good advice in that your plan has a 20% chance of going bust using your planned withdrawals. But WDR shows that by adjusting to current conditions and reducing spending temporarily, you have a chance to survive. It is all probabilities. You can't expect to get a "real" number for the outcome. This is not engineering where you can get the exact size of the required bridge beam if you can get the correct input data.
Don't expect too much from the forecasts. Other factors like global conflict, natural disasters, and your health in old age might be the most important factors. And they aren't considered in anyone's program.
Sorry, didn't mean to preach.
Q. Your reply is just great! I think I finally have a grasp on what forecasting a retirement accomplishes. You had to "preach" on the subject to get it across. I kept trying to determine which forecast provides the "best result." Now I think I understand that none provide that.
Still, I believe that your program provides the "best method" for forecasting that I have seen. Now I know that does not mean that it provides the best result. However, it provides the best method for manipulating variables that affect the outcome - variables that no one can predict accurately, whether they are historical or random based. The flexibility it provides allows more control over the selection of variables that determine the outcome. And since the whole process is a matter of chance anyway, it seems better to let the responsible party roll the dice.
And if I can be so presumptuous as to ask you what is your opinion on the viability of a Forecaster forecast that results in:
1) No errors on Calculate function,
2) 80% success on most Monte Carlo simulation runs, and
3) 99% success using WDR simulation runs.
A. The calculation phase is most useful for estate planning. It shows how all your planned gifts, premiums to insurance trusts, beneficiary instructions, etc, etc work. And the resulting cash flow is mapped out for you. But it assumes that your specified variables remain as specified over your life. There are no probabilities calculated for all the individual estate planning objectives. Of course, it does catch gross errors in financial planning.
Repeating my comment from before: Your results are giving you good advice in that MC shows your plan has an 80% chance of success using your planned withdrawals. But WDR shows that by adjusting to future conditions and reducing spending temporarily, you have a 99% chance to survive. Seems like WDR is demonstrating that common sense is always a good strategy. Plus it is giving you some sensible guidance for that budgeting.
As I said before, using the WDR is all new to me too. I have no guidance from the Klinger article for the real world case with taxes and IRAs, etc. So my suggestions here are really just guesses. That said, here are my suggestions:
Run the WDR simulation each year using all up to date numbers to arrive at a prudent withdrawal for the year. Have I mentioned that I am assuming that when you run the WDR, you are adjusting the WDR parameters and also your specified living expense amount to get an acceptable median plot line in the Yearly Expense Funding chart? This might require a reduction in the year's specified living expense amount. But this is WDR's way of telling you how to budget yourself based on the current conditions.
Q. Thanks again for your input on my comments about results, method and my case. I was actually "fishing" for your personal opinion of the viability of my forecast results, not only their appropriateness, but their satisfactoriness.
You write that using WDR is not clear about "...real world case with taxes and IRAs...” Since the "Starting value" and "Ending value" determined by all of the variables included in the Calculate function also appear on the MC input screen, aren't these real world variables accounted for in MC simulations? They do not appear on the WDR input screen, though, so does that mean they are not accounted for in those simulations? I can't believe how much there is to know about how these simulations work and thereby the understanding of the results.
A. I wrote that Klinger's paper is not clear about taxes and IRAs, etc, so there is no guidance there about real world usage. As for Forecaster's WDR functionality, it was designed with the goal of controlling the net funds available for expenses, not the portfolio withdrawals. The actual withdrawals vary depending on how much other income is available after taxes for your expenses.
My suggestions are my own interpretations of how to use this simulation, which I created using my own ideas about the effects of taxes and IRAs. You have as much or more experience than I do with trying to use it.
The "Starting value" and "Ending value" determined by all of the variables included in the Calculate function appear on the MC input screen only for your reference. Of course the starting value is used at the beginning of each MC simulation trial run. The ending value is just shown as a theoretical "Par" value, and is referenced in the resulting table below the MC histogram. This shows the probability of finishing with that ending value, but each MC simulation run finishes with a different ending value, which results in the probability distribution.
The "Starting value" and "Ending value" don't appear on the WDR input screen because there's not room for them, and besides they are not relevant to the data input required. But as with the MC, when running the WDR simulation, again the starting value is used at the beginning of each WDR simulation trial run. And again each WDR simulation run finishes with a different ending value, which results in the WDR probability distribution. The difference is that during the WDR simulation runs, the living expense target may be changed according to the WDR parameter values, and so the distribution of ending values here is different than the MC runs.
By the way, I'm interested in your final results with the WDR. What is the plus or minus percentage difference between your "First Year's Total Funds For Retirement Expenses" and your "Final Year's Median Value Of Funds For Retirement Expenses". These numbers are under the "Withdrawal Accounting" tab,
Q. Thanks for the further explanation of the MC and WRD simulations, and for informing me that they do use the beginning value, other income sources (I presume) and income goal. And I assume the simulations also account for taxes and IRA-type operations like MRDs. Also, thanks for noting the detail about ending value showing up as par in MC, which helps me understand how it fits in.
Here are my values as you requested: "Final years median value..." is 6.78% higher than "First years total funds...” I based the percentage gain on the first years total figure. According to your videos, this shows that the income goal has kept up with inflation by increasing 6.78% in real dollars, according to the COLA input.
As you probably know, the FIRECalc program achieves 100% success with the same input variables. Does the historical roll of the dice it uses always beat random number simulations?
A. Thanks for the numbers. And your assumptions are correct.
As for FIRECalc, I doubt that the historical roll of the dice it uses always beats random number simulations. A random number generator could theoretically generate any sequence of values. FIRECalc's results would directly depend on the fixed historical price data it was using. I have not spent any time with FIRECalc, so I can't comment beyond that.
I have enjoyed our conversation. As a developer, it always helps to write out in narrative form a description of the operation that is being designed. As a programmer, it is the last thing I want to do. But having a responsive listener makes it much more interesting.
I am thinking of using our e-mail conversation to create a help page for the Web site in the form of a Q and A session. I will edit your questions and my responses so that other new users may learn from our exchange. Most users probably won't write to ask these questions.