FIRECalc rewrite

[Rich_by_the_Bay]

Here's the article I mentioned above - it's about drawing down your bonds first in the decumulation phase.

[ERD50]

Quote:

Originally Posted by Rich_in_Tampa Here's the article I mentioned above - it's about drawing down your bonds first in the decumulation phase.

Rich, if I read that right, the results should not be surprising at all. It looks like he is using a totally static model. He just assumes stocks return the same higher real rates than do bonds. Same numbers every year.

In that case, sure it makes sense to spend down the bonds first. You want to give the highest return investment the longest amount of time to grow (other things being equal).

So it would be interesting to see what would happen with the data set in FireCalc.

-ERD50

[Rich_by_the_Bay]

Quote:

Originally Posted by ERD50 Rich, if I read that right, the results should not be surprising at all. It looks like he is using a totally static model. He just assumes stocks return the same higher real rates than do bonds. Same numbers every year. In that case, sure it makes sense to spend down the bonds first. You want to give the highest return investment the longest amount of time to grow (other things being equal). So it would be interesting to see what would happen with the data set in FireCalc.

Yup. Don't want to ambush Dory's thread, but I agree. However, for this to work, the asset types need to be held long enough to allow smoothing of volatility so, for example, 2 years of bond burn, then all stocks forever would not be as likely to survive as 7 or 10 or 15 years of fixed before tapping the stocks.

Boy would that be awesome to have within FC. Unfortunately, after bonds are burned you have additional choices: just start digging into stocks forever, or immediately convernt $xxx of stocks into bonds to start burning again (buckets of money approach).

[tightasadrum]

Along the same lines as Rich mentioned, how about,

1. more lines for reduced withdrawal in income page.

I have several investments in real estate. Unless I group then together as if I'm selling it all, I don't have needed flexibility. How complex does adding several more lines make the program? Four or five more lines would helpful.

2. The "how invested" page is ackward to me. How about putting down a table of all major catagories with input for % allocation. I'd like to see a catagory for REITS and commodity futures as well. Are TIPS on there now? Can't remember, but probably are.

[Cadence]

Dory,

Please consider using median values of portfolio results rather than averages on the results page. The distribution of portfolio returns over time is not a bell curve, it is a log normal distribution. Many scenarios will return less than the average, a few will return much larger than the average. In this case, the median is less, sometimes much less than the average. Therefore it is unlikely than any investors specific portfolio return will perform at or better than the average ending value of all scenarios.

If it were me I'd like to see the results page displayed something like:

50% probability of portfolio ending $900,00 or greater
80% probability of portfolio ending $500,00 or greater
90% probability of portfolio ending $100,00 or greater
95% probability of portfolio ending $50,00 or greater
etc.

It's my guess that its not understood that although a portfolio many have a 95% survivability rate, because of log normal distribution, a great many scenarios will move perilously close to 0.

I'll say again, I think firecalc is excellent. But I do believe that it gives users an unrealistic level of confidence.

Dory - I'm willing go into more detail on my thoughts on this if you are interested.

[jIMOh]

here's some ideas

1) allow entering of current ages, ER age, regular retirement age and late retirement age within one running of it, giving probale successes for each age range.

1a) alow different entries for spouses. me age X, ER age of X+20, R age of X+30, LR age of X+40. Wife aged Y, ER age of Y+7, R age of Y+10, R age of Y+13.

2) allow a person to enter account types (Roth, 401k, taxable accounts) and tax rates for each account (based on current tax tables). Give an advanced option which can calculate RMD's and compare to income needed.

2a) allow different entries for spouse (2 roth lines, 2 401k lines etc...)

3) allow Firecalc to remember me, so I don't have to enter all info when I go to it. I ran it once around 6 months ago... in a year or so, when I hit some type of milestone or lump sum, I'd like to rerun it, but just update the initial assumptions I entered.

3a) allow to remember spouse info

4) allow firecalc to use yearly returns BACKWARDS (so 2007 return, then 2006, then 2005, then 2004, then 2003) to test strategy using various other market conditions. I have done reading which suggest if the 2000-2002 bear came earlier in retirement, it takes most SWR and ratches them down a notch. Allow user to see what this does.

4a) give the user the ability to put a "bear market" the first X years of retirement which have an overall return of -y% and see what this does to overall success rates.

[dory36]

Quote:

Originally Posted by Cadence Dory, Dory - I'm willing go into more detail on my thoughts on this if you are interested.

Sure -- please do. I understand the confidence intervals, but elaboration will be useful when I start trying to compile all this into programming tasks.

dory36

PS - just by chance, you're not off a motoryacht named Cadence are you?

[Rustic23]

I never figured out how to do this in the present FireCalc, so it may be possible, but how about the death of a spouse? Pension check goes down by 40%, SS by a third. Would be helpful to be able t what if different dates in the future.

[Lusitan]

Thanks for all your work on FireCalc, Dory - it's a great tool to play with when tinkering with FIRE plans.

Others have made great suggestions, far more substantial than mine, but here's one thing that I think should be changed:

FIRECalc asks for your current portfolio value (X) and then it allows you to input (Y) additional savings per year until your retirement year ZZZZ.

When viewing results, FIRECalc always starts out by saying "Your plan is to spend $, which is W% of your starting portfolio ..."

But for the percentage, it calculates the percentage based on your CURRENT portfolio, not your portfolio upon reaching ER (which is the important one). This makes for some eye-popping "withdrawal rates" that really are out of place. The end result calculations are fine, but this withdrawal rate thing really sticks out.

Who cares what percentage of my current portfolio I will be withdrawling starting in 2020 when I retire ... what's important is what percentage of my 2020 portfolio I'll be withdrawling when I retire in 2020.

Thanks for taking a look at this and other changes!

[Lsbcal]

Another thought for the summary page. I use advanced FIRECalc and request the spreadsheet summary. Then I use the Excel functions to get the average, stddev, and min for the portfolio. I am particularly interested in the min (maximum drawdown) on the portfolio because I know I won't have the guts to live with an excessive drawdown. Also I want to know the year the max drawdown occurs to see if it's at the end of the test period or somewhere in the middle.

Currently the a historgram is showing the portfolio end value is shown. To complement this, it would be nice to see a histogram of the (worst) drawdowns for each starting retirement sequence.

As an alternative a few values could be produced:
average worst case drawdown =
(this is the average of the minimums for each sequence)
worst drawdown =
year of worst drawdown =

Les

[dory36]

Lots of excellent suggestions. Keep 'em coming. The "feature freeze" will probably be in about 2-3 weeks, when I move back onto the boat for a while.

To get a notice when there is a test version to play with, sign up for the mail list at http://firecalc.com/newfc.php.

dory36

[dory36]

An idea ... please give me some feedback.

The problem: long term retirements, say 50 years, miss the problem years of the mid 1960s, since there are no 50 year data sets that start with the rough years of the mid 1960s. (100 year retirements, if you believe in extreme life extension possbilities, miss the Great Depression.)

The hypothesis: Failures occur when there are down years in the early part of the retirement cycle. If the early years are OK or better, no problem. That is anecdotal, but I think the data will support it. So, we can predict the long term outcome based on the first ~5 years. If this hypothesis is correct, we can include data all the way up to 2002 for prediction of retirement plan success of unlimited retirement terms.

Thoughts? It makes me wish I had been storing all the gazillion runs, so I would have a large body of data for analysis. Statisticians... how can I come up with a rule, such as "If the 5th year portfolio balance is below 63.3% of the starting balance, then ..."?

Whatever year seems most valid (the earlier the better), I need to develop a mapping of Nth year ratio to later year outcomes.

dory36

[Cb]

Quote:

Originally Posted by dory36 Lots of excellent suggestions. Keep 'em coming. The "feature freeze" will probably be in about 2-3 weeks, when I move back onto the boat for a while. To get a notice when there is a test version to play with, sign up for the mail list at http://firecalc.com/newfc.php. dory36

That link doesn't work for me.

Cb

[ERD50]

Quote:

Originally Posted by dory36 An idea ... please give me some feedback. The problem: long term retirements, say 50 years, miss the problem years of the mid 1960s, Whatever year seems most valid (the earlier the better), I need to develop a mapping of Nth year ratio to later year outcomes. dory36

Glad you are going to try to tackle this. I'm not sure there is any good answer. But, a suggestion:

Maybe this would take too many calculations, but what if you wrapped around and just kept tacking on the 'extra' years from the start of the data up to where you start to overlap.

Example: 50 year run cannot currently utilize data past 1956 to get a full 50 years. So do the normal runs, then when you run from 1957 to 2006, use each year from 1871 to 1956 tacked on to complete the data set. Then run 1958 to 2006, and tack on two year periods from 1871 to 1958, etc, etc, etc.

Too much computer time?

Off hand, that seems like a very negative way to look at it, you could have two of the worst sequences to deal with in those 50 years. But, let's see... 1974, 1929 - just 45 years apart! Maybe two disasters should be 'expected' over a 50 year period?

That's a grim thought to end a holiday on!

-ERD50

[ERD50]

Quote:

Originally Posted by Cb That link doesn't work for me. Cb

it is not formatted correctly,

just paste this into a new window and it went through for me:

firecalc.com/newfc.php

or try one of these:

FIRECalc

corrected version of link

-ERD50

[twaddle]

Quote:

Originally Posted by dory36 The problem: long term retirements, say 50 years, miss the problem years of the mid 1960s, since there are no 50 year data sets that start with the rough years of the mid 1960s. (100 year retirements, if you believe in extreme life extension possbilities, miss the Great Depression.) The hypothesis: Failures occur when there are down years in the early part of the retirement cycle. If the early years are OK or better, no problem. That is anecdotal, but I think the data will support it. So, we can predict the long term outcome based on the first ~5 years. If this hypothesis is correct, we can include data all the way up to 2002 for prediction of retirement plan success of unlimited retirement terms.

Predictions of success? Correlation between the start of the sequence and failure? Err, sounds like a slipery slope to me. Definitely not in the spirit of the old FIREcalc.

But it would be an interesting alternative to FIREcalc. My guess is that the new FIREcalc would conclude that anybody who retired in the year 2000 would be doomed to failure (i.e., negative real returns at the end of 5 years).

[jay8]

How about more asset classes for back-testing:


1. Mid-cap
2. Emerging market
3. International Large Growth
4. International Large Value
5. REIT
6. TIPS
7. Intermediate US Treasury

[dory36]

Quote:

Originally Posted by twaddle Predictions of success? Correlation between the start of the sequence and failure? Err, sounds like a slipery slope to me. Definitely not in the spirit of the old FIREcalc.

Heh... I'm not thinking about substituting this sort of thing for any sequence where there are data.

Rather, once we run out of data (starting after 1967 for a 40 year term for example), a separate result would be reported, something like:

"We can't look at historical results for a 40 year term starting after 1967, but we can compare starting trends. For your own scenarios, 66% of the runs eventually failed, when the portfolio after 5 years was below 57% of the starting point. We might expect that same ratio would apply to the 35 additional starting years for which we don't have a full 40 years of data. Of the additional 35 starting years we can test this way, 6 of the starting years resulted in 5th year portfolio balances that were at or below 67% of the starting amount. Based on these trends in the partial sequences, we might expect that 4 of the 35 partial sequences would eventually result in a failed portfolio. This would change your overall success rate to 83%."

dory36, never using a single word when 10 will do...

[dory36]

Quote:

Originally Posted by ERD50 Maybe this would take too many calculations, but what if you wrapped around and just kept tacking on the 'extra' years from the start of the data up to where you start to overlap.

I don't think it would be reasonable to assume that any specific year would be a plausible subsequent year to 2007, for "splicing" data to make longer sequences.

If we take the ending balance of each short sequence, and do a completely new run using that ending balance as the starting portfolio, and set the term to be the number of years that we are short, then we could come up with a probability that each short sequence would succeed or fail.

Each of the full sequences either succeeds or fails. Each of the partial sequences would have a percentage of likely success.

But I suspect the impact of making every run do ~40 times the number calculations as it currently does might make this approach too demanding for the server.

[ERD50]

Quote:

Originally Posted by dory36 I don't think it would be reasonable to assume that any specific year would be a plausible subsequent year to 2007, for "splicing" data to make longer sequences.

I tend to agree, but what other choice is there if you want to lengthen the data stream (other than a statistical approach)?

Quote:

If we take the ending balance of each short sequence, and do a completely new run using that ending balance as the starting portfolio, and set the term to be the number of years that we are short, then we could come up with a probability that each short sequence would succeed or fail.

Isn't that the same thing?

Quote:

Each of the full sequences either succeeds or fails. Each of the partial sequences would have a percentage of likely success. But I suspect the impact of making every run do ~40 times the number calculations

It might not be all that data intensive with your approach. Assuming most people shoot for high % success rates (95% or greater typically):

Call the 'full data' years 'seq A', the 'remaining years 'seq B'.


1) 'Failures' in 'seq A' are failures. Do not extend those runs.

2) Starting with the worst of the runs of 'seq A', run against 'seq B' as you suggest.

3) If you get any failures in 'seq B' with that starting value, go to step 2 and run again with the next lowest balance for 'seq A'.

4) Repeat steps 2-4 until you get a success for that balance against all 'seq B' runs.

What you would be giving up is complete data on average ending balances, max and min. But you would still have a success rate and you could still seek out a starting portfolio or spending rate for X% success.(edit/add) It might take a few more iterations, you don't know ahead of time if you will have some or no failures in the 'Seq B'. If 'Seq B' failures take you far from your XX% success rate, you might need to go back and bump up portfolio or decrease spending to improve success rate.

I think a disclaimer to note that the parametric data is not complete for those longer runs is a reasonable compromise. IMO, the parametric data is interesting, but a far lower usefulness than the success/failure number.

-ERD50