FIRECalc throws away/ignores new data

[yachi]

I guess I'm not really asking a question, just venting and warning others.  Like a prudent early retiree, I've been putting in a long 58-year period into FIRECalc.  It spits out 1 failed scenario out of 93 which sounds great, but look what happens when I play with the length of time:
58 years: 1 failed scenario of 93
55 years: 3 failed scenarios of 96
50 years: 5 failed scenarios of 101
45 years: 4 failed scenarios of 106
40 years: 2 failed scenarios of 111
30 years: 0 failed scenarios of 121

It makes no sense for my 'stache to have a 95% chance of lasting 50 years, and a larger 98.9% chance of lasting 58 years but that's what FIREcalc happily spits out.  The reason this is happening is because Firecalc throws away periods that are shorter than what you model, so periods starting in 1964 are not included in my 58-year model.  That's shortsighted because, as you can see above, you can tell some of those scenarios are failures.  Failures won't recuperate no matter how much time you give them.  You can't tell if the less than 58-year periods are successes, because you don't have the entire period so they could still fail, but you can count the failures.

[FIPurpose]

That's part of the reason why there isn't a strong reason to do any better than 80%. When looking at stuff on FIREcalc, a 85%, 90%, 95%, and 100% success rate should all be viewed with about the same amount of assurance. So don't worry too much about FIREcalc having to drop a year of data or 2. It doesn't change the outcome significantly.

[Metalcat]

There are so many variables that FIREcalc can't even account for that numbers at a certain point are totally meaningless.

One person's 80% might be way more secure than another person's 100% depending on the assumptions they input.

Also, FIREcalc isn't useful for predicting the future, it's useful for making decisions as to how to handle your finances in the present.

Basically, you shouldn't stress it because you also shouldn't be using it for reassurance.

[Simpli-Fi]

I like the way they explain it

“Consider an analogy: Suppose you are building a house in Honolulu. No one could predict the temperature for any given future date during the decades the house will be used. But if you know that it has never been under 52° in that location in all of recorded history, you could make an intelligent judgment about how much heating capacity is enough.

Planning for an Anchorage-style winter would be a true waste of money that could be better used elsewhere.”

[RWD]

Isn't this because there are less scenarios available in the data as you increase the time period? You can only fit so many scenarios in the number of years we have data for. So there are no 58-year periods which start in the year 1966 (55 years ago) yet which was the worst possible year to retire.

This is why for a statistical model like this with limited data you should stick with 30 year time periods to get more sample periods. Also, consider using cFIREsim as an additional simulator.

[maizefolk]

Yes, this is an inherent issue with doing the calculation over long time frames, it's not anything specific to FIRECalc. But it means that increasing the length of your projected retirement can DECREASE rather than increase the risk of failure, which doesn't make intuitive sense.

I think we could get around this by using the same math used for calculating life expectancy to calculate portfolio survival, but so far that approach hasn't really caught on.

Things to know that mean my projections may not line up with the standard ones you might get out of a website like cfiresim: 1) I use monthly data on stock returns (shiller data) to calculate a lot more total scenarios. 2) that means I also calculate withdrawals on a monthly basis, not one lump sum per year 3) to calculate failure rates I use portfolio life expectancy, which I think provides a more accurate estimate of failure rates over extremely long retirements than looking at the number of failures out of all the time intervals that are as long as your estimate retirement in historical data (traditional Trinity approach).



My approach is the green line, the traditional approach is the blue line. Intuitively, failure rates shouldn't decrease as we go to longer retirement lengths, but the problem is that really bad years (like the mid 60s) start dropping out of your dataset once your retirement window gets long enough.

[yachi]

Crap, I had a good reply going to many of you, but refreshed the page and lost it.

Anyway, I think it is a very useful upgrade, almost a necessity, for FIRECalc to look at partial time periods to see if they are already known failures.  So if the time period starting in 1966, resulted in a failure in 2006, it would also be counted as a failure for a period that would end in 2026.  The percentage can still be based on the number of actual finished 58-year periods, but knowing the number of failures in shorter periods is actually quite useful.

I'm not sure if CFIREsim splits the data up the same way or not, it's been a while since I used it.  I do like that you can look at individual start/stop time periods in the CFIREsim graph.

[Villanelle]

To me, this all goes back to what one should use FIREcalc (and all the similar models/"planners" for).

It just gives me a rough idea of what's likely.  Since I know that if we are in sustained significant down markets or stagflation, I will make cuts to my budget (and my budget has enough fat that it will be easy to trim some) and/or find a little bit of income, I already don't need perfection.  If one scenario is 87 and another is 93%, it is basically saying the same thing for me--'this is likely going to work perfectly, and if not perfectly then I'll have to enact that "minor cuts during bad times" scenarios I already have in mind, that that will work just fine, too.'  If there's a 3% higher chance that I have to make a few minor cuts in one scenario/model/projection/choice about how to handle the numbers, I'm still very, very good with that.  Because not being able to vacation out of state for a couple years in retirement is > not being able to vacation out of state for a couple years because I am still *working full time*.

FIREcalc never purported to be a razor-honed exact prediction of anything.  It's just a perspective-giver.  And to me, the numbers from that OP all give the same perspective, which is... barring black swan shit and/or total unwillingness to flex with life, this is going to be just fine.

[Metalcat]

To me, this all goes back to what one should use FIREcalc (and all the similar models/"planners" for).

It just gives me a rough idea of what's likely.  Since I know that if we are in sustained significant down markets or stagflation, I will make cuts to my budget (and my budget has enough fat that it will be easy to trim some) and/or find a little bit of income, I already don't need perfection.  If one scenario is 87 and another is 93%, it is basically saying the same thing for me--'this is likely going to work perfectly, and if not perfectly then I'll have to enact that "minor cuts during bad times" scenarios I already have in mind, that that will work just fine, too.'  If there's a 3% higher chance that I have to make a few minor cuts in one scenario/model/projection/choice about how to handle the numbers, I'm still very, very good with that.  Because not being able to vacation out of state for a couple years in retirement is > not being able to vacation out of state for a couple years because I am still *working full time*.

FIREcalc never purported to be a razor-honed exact prediction of anything.  It's just a perspective-giver.  And to me, the numbers from that OP all give the same perspective, which is... barring black swan shit and/or total unwillingness to flex with life, this is going to be just fine.

Exactly.

It's so critical to distinguish what these outputs actually mean for.people in terms of their tangible life decisions.