Author Topic: firecalc and cFIREsim both lie?  (Read 97248 times)

sol

  • Walrus Stache
  • *******
  • Posts: 8492
  • Age: 42
  • Location: Pacific Northwest
firecalc and cFIREsim both lie?
« on: April 09, 2014, 05:40:18 PM »
I've been playing with firecalc and it's mustache-forum-grown technical equivalent cFIREsim for quite a while, and I think I've identified a conceptual problem with the approach.  Please correct me if I'm confused about this.
 
These tools make predictions about the long term success of your hypothesized withdrawal strategy based on the number of historical periods in the market record that correspond to your time horizon.  So if you ask for a 30 year projection, it will look at every 30 year period in the market's history and then tell you how many of those 30 year periods held up under your proposed withdrawals.
 
The problem is that these historical periods are necessarily biased by recent returns.  For example, the stock market has more than doubled in the past five years so if you had asked for a 30 year projection five years ago and asked for the same projection today, you'd get very different answers because the more recent simulation contains five additional years of incredibly profitable gains.
 
And this is exactly backwards from what we want, as retirement prognosticators.  Five years ago the market was depressed below long term averages and today the market is (maybe) a bit overpriced, meaning that for equal sized nest eggs your chances of FIRE success were higher five years ago than they are today.  But instead of recognizing this, both of these tools give you the exact opposite answer, than you're better off retiring today than you were five years ago.
 
From one perspective, the last five years of 20%+ per year market returns actually REDUCE the forecasted rate of return while INCREASING the firecalc/cFIREsim success projections (by adding additional positive years to the period of record).
 
This isn't a flaw in the way the tools are built or executed, it's a flaw in the conceptual model of how the theory behind them works.  Rapid market runups artificially inflate the success predictions while reducing return expectations in the real world.  They make things look rosier when instead they should be looking a bit darker.
 
Is it a huge difference?  I'll let others opine on that but given that most people here seem to be using these tools to get between 90% success and 95% success ratios, a few percent here or there could be significant to them.
 
This is the exact same problem another forum member mentioned here recently, about how everyone who is invested in the market is much more likely to hit their "target number" in the middle of a market surge than during a downturn, which virtually guarantees that we're all going to think we're retirement eligible based on overvalued assets.
 
My proposed solution?  I think I'm going to look at forecasting my retirement based on our assets and the current P/E ratio, rather than just assets alone.  Having a million dollars at the height of the tech bubble is a lot different from having a million dollars at the depths of the recession, and the P/E should help clarify where in that cycle we currently sit.
 
 
 

beltim

  • Magnum Stache
  • ******
  • Posts: 2841
Re: firecalc and cFIREsim both lie?
« Reply #1 on: April 09, 2014, 05:47:12 PM »
This is a really, really good question, with a lot of content to chew on.  Two quick thoughts, though:

1) If you think of the stock market returns we've experience as being a sample of the universe of potential stock market returns, then each additional year gives you better sampling of the "true" average.  Better sampling will reduce the error in your model.  Right?

2) Rather than use a simple PE, I'd use something with a longer history.  The Schiller PE10 has much better predictive value than the last 12 months PE, for example.

warfreak2

  • Handlebar Stache
  • *****
  • Posts: 1140
  • Location: UK
    • Music by me
Re: firecalc and cFIREsim both lie?
« Reply #2 on: April 09, 2014, 05:53:08 PM »
The problem is that these historical periods are necessarily biased by recent returns.  For example, the stock market has more than doubled in the past five years so if you had asked for a 30 year projection five years ago and asked for the same projection today, you'd get very different answers because the more recent simulation contains five additional years of incredibly profitable gains.
I don't see why they'd be biased towards recent years; it just happens to have that data in there, and uses it the same as the rest of the data. If anything, it should be weighted against recent years: for example, if 2013 is the last year in the database, then only one of the 30 year periods includes that data (1984-2013 inclusive), while the year 1970 is included in all 30 of the 30-year periods starting between 1941 and 1970 inclusive.

Quote
But instead of recognizing this, both of these tools give you the exact opposite answer, than you're better off retiring today than you were five years ago.
Ah, I think that's a misunderstanding. It doesn't mean you'd be better off retiring today than retiring five years ago; it means it now thinks you would be better off retiring at all (whenever you do retire), than it used to think five years ago.

Unless you're trying to time the market, the assumption is that whenever you retire, you have the same chance of success. That chance of success will be measured more accurately if it's based on more data. If including more data made it less accurate, we could get better answers by deleting the data!

arebelspy

  • Administrator
  • Senior Mustachian
  • *****
  • Posts: 28025
  • Age: -999
  • Location: Seattle, WA
Re: firecalc and cFIREsim both lie?
« Reply #3 on: April 09, 2014, 05:58:37 PM »
Yes, after a bull market cFIREsim will have a few extra simulations of success, which may lead to portfolio failure down the road (though I'm skeptical - anyone with enough brains and foresight to be planning out their retirement via a tool like cFIREsim will make the necessary judgments and changes).

OTOH, after a bear market cFIREsim may be under representing how safe your portfolio is, and you might work even longer than necessary.

Further, I don't personally believe a few percent will make or break a retirement.  So there was the same level of historical failures in the past as if you had run FIRECalc in 2009, but running it now has a few higher successes so a bit higher percent, and that means yours is more likely to fail now?

Maybe someone with the time and cleverness can see if you can do something neat with P/E or Schiller P10 and cFIREsim (i.e. compare success rates of years in which it was above or below X).
We are two former teachers who accumulated a bunch of real estate, retired at 29, spent some time traveling the world full time and are now settled with two kids.
If you want to know more about us, or how we did that, or see lots of pictures, this Business Insider profile tells our story pretty well.
We (rarely) blog at AdventuringAlong.com. Check out our Now page to see what we're up to currently.

beltim

  • Magnum Stache
  • ******
  • Posts: 2841
Re: firecalc and cFIREsim both lie?
« Reply #4 on: April 09, 2014, 06:00:13 PM »
Also, I'd suggest another kind of retirement calculator, specifically one that uses Monte Carlo simulations using historical data.  Vanguard has a free one:
https://retirementplans.vanguard.com/VGApp/pe/pubeducation/calculators/RetirementNestEggCalc.jsf

FIPurpose

  • Handlebar Stache
  • *****
  • Posts: 1047
  • Location: WA
    • FI With Purpose
Re: firecalc and cFIREsim both lie?
« Reply #5 on: April 09, 2014, 06:00:18 PM »
I believe one other major point is that this ignores the fact that most people do not hold 100% total stock market. There is some amount of balance between stocks/bonds. That means that overvalued/undervalued stocks at particular periods of time provide steadier, better, and more consistent returns.

EDIT: Provided that you are regularly rebalancing your portfolio.
« Last Edit: April 09, 2014, 06:02:10 PM by flyingcircle »

mulescent

  • Stubble
  • **
  • Posts: 114
Re: firecalc and cFIREsim both lie?
« Reply #6 on: April 09, 2014, 06:04:59 PM »
That is an interesting question.  I see two potential problems, and maybe formalizing them a bit will help.  First, there is the fact that the most recent five years are really great years.  They'll be included in the 30 year periods which are being drawn randomly and, being really good, they'll bias the results.  I'm not so worried about that, since there are over a hundred 30 year periods to draw from.

Another potential problem, which I think you raised and also concerns me is the inherent assumption that the starting point for the simulation (your NW today) will not influence the likelihood of any of the subsequent outcomes.  If you argue that we are near/at a market peak, then it's entirely reasonable to believe that what happens next will not be a randomly chosen 30 year period but one that tends to come after big market run-ups (i.e. a downturn).  Your idea of modifying the algorithm to take P/E into account is a reasonable one.  You could restrict the pool of 30 year windows to those occurring some short time after similar P/E values.  An even better algorithm would try to figure out whether that P/E was hit on the way up or down, so that you could account for momentum, too.

warfreak2

  • Handlebar Stache
  • *****
  • Posts: 1140
  • Location: UK
    • Music by me
Re: firecalc and cFIREsim both lie?
« Reply #7 on: April 09, 2014, 06:11:31 PM »
there is the fact that the most recent five years are really great years.  They'll be included in the 30 year periods which are being drawn randomly and, being really good, they'll bias the results.
No more than including the 6s when you roll dice biases the results by making them, on average, higher. It's not bias if they have equal weighting, and certainly not bias if they have significantly lower weighting due to appearing in far fewer of the 30-year spans.

beltim

  • Magnum Stache
  • ******
  • Posts: 2841
Re: firecalc and cFIREsim both lie?
« Reply #8 on: April 09, 2014, 06:12:53 PM »
The times a 4% withdrawal rate has failed according to cFIREsim and the Shiller PE on Jan 1 of that year:
1906: 20.1
1965: 23.27
1966: 24.06
1967: 20.43
1968: 21.51
1969:21.19
1973: 18.71

The average PE10 all-time is 16.52, so these are all at least 10% higher than the average.  The 4% rule has failed in 7/39 cases where the Shiller PE was above the average, and 6/15 times that the starting PE10 was above 20.

arebelspy

  • Administrator
  • Senior Mustachian
  • *****
  • Posts: 28025
  • Age: -999
  • Location: Seattle, WA
Re: firecalc and cFIREsim both lie?
« Reply #9 on: April 09, 2014, 06:20:39 PM »
The times a 4% withdrawal rate has failed according to cFIREsim and the Shiller PE on Jan 1 of that year:
1906: 20.1
1965: 23.27
1966: 24.06
1967: 20.43
1968: 21.51
1969:21.19
1973: 18.71

The average PE10 all-time is 16.52, so these are all at least 10% higher than the average.  The 4% rule has failed in 7/39 cases where the Shiller PE was above the average, and 6/15 times that the starting PE10 was above 20.

That's good data, thanks!
We are two former teachers who accumulated a bunch of real estate, retired at 29, spent some time traveling the world full time and are now settled with two kids.
If you want to know more about us, or how we did that, or see lots of pictures, this Business Insider profile tells our story pretty well.
We (rarely) blog at AdventuringAlong.com. Check out our Now page to see what we're up to currently.

warfreak2

  • Handlebar Stache
  • *****
  • Posts: 1140
  • Location: UK
    • Music by me
Re: firecalc and cFIREsim both lie?
« Reply #10 on: April 09, 2014, 06:47:27 PM »
Most of those years are very close together, which demonstrates how the years in the middle are weighted more by appearing in more 30-year windows. It's probably the same crash which causes ruin in each of those 30-year windows, so that one crash appears 5-6 times in the list.

sol

  • Walrus Stache
  • *******
  • Posts: 8492
  • Age: 42
  • Location: Pacific Northwest
Re: firecalc and cFIREsim both lie?
« Reply #11 on: April 09, 2014, 06:49:11 PM »
1) If you think of the stock market returns we've experience as being a sample of the universe of potential stock market returns, then each additional year gives you better sampling of the "true" average.  Better sampling will reduce the error in your model.  Right?

I think I'd be more inclined to believe that if the long term average was stationary over time.  Except that no stock market in the history of the world has ever had a consistent long term average over hundreds of years, because stock markets reflect economies and economies evolve.  If the period of 2014 to 2044 averaged 1% return, would you believe that the Dow was due for a couple of huge decades to get us back to our long term average, or would you believe that the US economy had stalled?

I don't see why they'd be biased towards recent years

Not biased towards, just biased by.  The long term CAGR of the US market is about 0.2% different through 2007 than it is through 2013.  0.2% is a small number unless you're talking about compounding market history, in which case it is approximately a 100% difference in dollar value.

Yes, after a bull market cFIREsim will have a few extra simulations of success... OTOH, after a bear market cFIREsim may be under representing how safe your portfolio is, and you might work even longer than necessary.

Right, this was exactly my point.  If there have been recent dramatic swings in the market, then the tools will give you different answers and they will be different in the way that is exactly backwards from reality.  Use it after a bear and you've worked too much, use it after a bull and you quit too soon.

Quote
Maybe someone with the time and cleverness can see if you can do something neat with P/E or Schiller P10 and cFIREsim (i.e. compare success rates of years in which it was above or below X).

Several people have mentioned this idea already, and I had actually already formulated a post on this very topic several months back and just never got around to posting it.  I dug it up, and have pasted it below.  It's a little rough, this was a first draft, but I think it highlights how using the CAPE is very relevant to predicting future success ratios.  The 1973 example beltim just posted as one of the 4% SWR failures is, notably, a lower value than we're seeing today, suggesting that a 4% SWR may not be safe at this moment in time.  This is exactly what I mean by needing to use P/E to make retirement predictions, not just the historical simulators.

...

Firecalc suggests that the returns in the few years immediately following your retirement are most significantly determinitive of your long term success rate at fixed spending amounts.
 
For example, for a 35 year retirement at 30k on 750k saved, firecalc suggests a 93% success rate.
 
But if the market has a 10% "correction" in the first year of your retirement, now you're looking at having withdrawn the 30k and then seen the remaining 720 reduced to 645. Re-evaluating the odds on 645k for the remaining 34 years shows a 72.5% success rate.
 
And this effect is not symmetric.  If the market surges 10% in the first year of your retirement instead of dropping then your odds do increase but only by about 4%.
 
To review, if the market goes up 10% in your first year your odds increase 4% but if the market goes down 10% in your first then your odds decrease by 20%.  Very lopsided.  Depending on your assumptions about the market's future, you can construct a scenario in which your EV on firecalc success in the future should drop on the day you retire regardless of when that is.
 
I think this effect can be partially comensated by using CAPE or P/E ratio or something else.  The 1973/4/5 example from the firecalc homepage shows dramatically different outcomes for the same nest egg and withdrawal combination in three consecutive years.  In 1973 30k on 750k goes bankrupt in 19 years, but P/E was ~18.  Same numbers in 1974 does okay, but P/E was ~12.  1975 with same numbers does fantastic, but P/E was only 8.3.
 
Here's the frightening part:  Today's P/E (trailing 12 month as of Feb 5 2014) is 18.56, higher than 1973.  Of course this ignores the importance of inflation in the 1970s, but I think it's still relevant.  Using the Cyclically Adjusted P/E should account for this effect, but in this example it doesn't make much of a difference (1973/4/5 are 18.7, 13.5, 8.9).

brewer12345

  • Handlebar Stache
  • *****
  • Posts: 1386
Re: firecalc and cFIREsim both lie?
« Reply #12 on: April 09, 2014, 06:56:43 PM »
Has hocus made an appearance on these boards yet?

BigRed

  • Stubble
  • **
  • Posts: 200
  • Age: 43
  • Location: NJ
Re: firecalc and cFIREsim both lie?
« Reply #13 on: April 09, 2014, 07:09:05 PM »
I recall reading an analysis by Pfau, I think, that did account for this effect, namely that the early years are extremely important to plan success and that recent returns are somewhat predictive of near future returns.

But, to some extent, you are engaging in the gamblers fallacy here.  If the market were truly correlated in the way you describe, the efficient market hypothesis would be roadkill.  While the EMH has some problems, the idea that a market fall is more likely at the present time ought to be priced into the current market value.  So, the best assumption from a naiive investor is that the market should still be expected to rise, and at the same rate as before.

sol

  • Walrus Stache
  • *******
  • Posts: 8492
  • Age: 42
  • Location: Pacific Northwest
Re: firecalc and cFIREsim both lie?
« Reply #14 on: April 09, 2014, 07:28:26 PM »
While the EMH has some problems, the idea that a market fall is more likely at the present time ought to be priced into the current market value.  So, the best assumption from a naiive investor is that the market should still be expected to rise, and at the same rate as before.

That's a totally sound theory, except that nobody here actually believes it.

I might be safe in saying that everyone on these boards recognized that the 2009 crisis was a buying opportunity. 

I realize that predicting future returns based on the CAPE is just a very simplified form of long term market timing.  I'm okay with that, because the long term data back it up.

BigRed

  • Stubble
  • **
  • Posts: 200
  • Age: 43
  • Location: NJ
Re: firecalc and cFIREsim both lie?
« Reply #15 on: April 09, 2014, 07:35:19 PM »
Ouch.  I suppose I earned that.  I don't believe in the EMH at all in the way my post implies, as I reread it now. 

The market timing implications of your theory seem like something I'd be hard pressed to act on, though.  What are the practical implications of your thinking?

arebelspy

  • Administrator
  • Senior Mustachian
  • *****
  • Posts: 28025
  • Age: -999
  • Location: Seattle, WA
Re: firecalc and cFIREsim both lie?
« Reply #16 on: April 09, 2014, 07:42:27 PM »
Has hocus made an appearance on these boards yet?

No.  Though I do recall a post by you a year or two ago asking that and saying something to the effect that a forum hadn't "made it" until he made an appearance.

Guess we've slid under his radar, for now.  ;)

(Those of you curious will have some fun googling.)
We are two former teachers who accumulated a bunch of real estate, retired at 29, spent some time traveling the world full time and are now settled with two kids.
If you want to know more about us, or how we did that, or see lots of pictures, this Business Insider profile tells our story pretty well.
We (rarely) blog at AdventuringAlong.com. Check out our Now page to see what we're up to currently.

brewer12345

  • Handlebar Stache
  • *****
  • Posts: 1386
Re: firecalc and cFIREsim both lie?
« Reply #17 on: April 09, 2014, 07:47:43 PM »
Has hocus made an appearance on these boards yet?

No.  Though I do recall a post by you a year or two ago asking that and saying something to the effect that a forum hadn't "made it" until he made an appearance.

Guess we've slid under his radar, for now.  ;)

(Those of you curious will have some fun googling.)

If we are all really, really lucky that troll will stay under his bridge. 

I just get a little nervous when I see such nice bait laid out for him.

RootofGood

  • Handlebar Stache
  • *****
  • Posts: 1361
  • Age: 39
  • Location: North Carolina
  • Retired at age 33. 5 years in, still loving it!
    • Root of Good
Re: firecalc and cFIREsim both lie?
« Reply #18 on: April 09, 2014, 08:02:06 PM »
OP, I don't think it matters whether the last 5 years are "good" or "very very good".  It just made the time series starting in 1980, 81, 82, 83, and 84 just a little bit better (well, a lot better) for terminal portfolio values.  But those 30 year periods starting in 1980-84 weren't at risk of failure at a 4% SWR anyway.

I don't see a problem with factoring in the market's recent run up over the last 5 years.  I'd feel better retiring with a million dollars in 2009 than I would with a million today, for example. 

wtjbatman

  • Handlebar Stache
  • *****
  • Posts: 1313
  • Age: 36
  • Location: Missouri
Re: firecalc and cFIREsim both lie?
« Reply #19 on: April 10, 2014, 06:51:04 AM »
Has hocus made an appearance on these boards yet?

No.  Though I do recall a post by you a year or two ago asking that and saying something to the effect that a forum hadn't "made it" until he made an appearance.

Guess we've slid under his radar, for now.  ;)

(Those of you curious will have some fun googling.)

Indeed I did (content at link is 5+ years old, but is still hilarious)

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #20 on: April 10, 2014, 06:54:03 AM »
Interesting topic.

I think that an individual who is close to FIRE and has enough education and experience under their belt should be able to pick up on the risks inherent in FIRE. It is already known that a market drop at the beginning of FIRE is a risk which cuts your success rate. And of course it is a (I don't want to use the word flaw or a lie, maybe downside?) downside to historical data usage that more successful years inflate your success rate while more low years inflate your failure rate.

I guess we just have to be those educated and experienced people. :)

As usual we can't take all the advice from the boards on blind faith. These are things we have to grok.

simonsez

  • Pencil Stache
  • ****
  • Posts: 796
  • Age: 33
Re: firecalc and cFIREsim both lie?
« Reply #21 on: April 10, 2014, 07:17:35 AM »
I apologize if this drifts or I am not understanding a concept but I have a question about firecalc and cfiresim.  If the assumption is that past does not predict the future but does give a reasonable range of what can happen given enough periods of time, why are the 30 year (or whatever number of years) periods ordered?

e.g. Results spit out 113 lines representing the 113 30 year periods but shouldn't we be seeing 113! (factorial) permutations  of 30 year periods if they are independent?  What does it matter if the random 30 year set includes 1871, 1915, 2009, and so on together?  They're supposed to be unrelated.

If the setup of these two websites (and others like it) did have all the permutations of n-year periods rather than chunks that move together through time, what the markets did in the last 5 years truly would not matter or bias the results, no?

mrigney

  • Stubble
  • **
  • Posts: 157
    • Running of the Fools
Re: firecalc and cFIREsim both lie?
« Reply #22 on: April 10, 2014, 07:30:08 AM »
From my longtime interpretation on the utility of firecalc/cfiresim, I think the software's main purpose is different than is being represented. Both pieces of software are great for a lot of things, but neither is a forecast of whether you can retire. Instead it's a historical guide; a bunch of Monte Carlo runs of a simulation; a probabilistic tool for your tool belt. But it is not (and I don't think it claims to be) a tool that gives you the probability that your retirement will fail or succeed.

A lot of my work has been done in ensemble numerical weather modeling, where you multiple (20, 30, 40, 50) runs to find a potential range of forecasts. However, if I run my model for historical weather events and tweak my model so that parameters fit the historical events well (analogously, firecalc/cfiresim is a perfect model of history b/c it's not a model...it is straight historical data), that has very little worth in persuading people that my model will now predict future events well. You simply can't rely on a model designed to fit historical data--or in the case of firecalc/cfiresim pure historical data--to have any real value as a forecasting tool. It is a good way of summarizing historical information that I think is extremely useful in making an informed decision about whether or not you can FIRE, but it is not in and of itself a forecast.

Now, with that said, if you take historical data and models and start incorporating P/E ratios and other relevant data...now you're starting to create a model that might have some predictive value. But fundamentally, your predictive model at that point is largely in what you decide to do with PE ratios, CAPE, etc...your model is still taking historical data from firecalc (success rate from firecalc/cfiresim is still just that...historical data condensed down to a percentage) and then you are making predictions based on current P/E, CAPE, etc.

So, I whole heartedly agree that these tools aren't a good forecast...they're not designed to be a forecast at all. Instead they're a historical Monte Carlo/ensemble run that gives you a whole lot of information that you can then make a forecast off of. Does that make any sense at all?
« Last Edit: April 10, 2014, 08:01:41 AM by mrigney »

nereo

  • Senior Mustachian
  • ********
  • Posts: 10800
  • Location: Just south of Canada
    • Here's how you can support science today:
Re: firecalc and cFIREsim both lie?
« Reply #23 on: April 10, 2014, 07:51:34 AM »
I apologize if this drifts or I am not understanding a concept but I have a question about firecalc and cfiresim.  If the assumption is that past does not predict the future but does give a reasonable range of what can happen given enough periods of time, why are the 30 year (or whatever number of years) periods ordered?

e.g. Results spit out 113 lines representing the 113 30 year periods but shouldn't we be seeing 113! (factorial) permutations  of 30 year periods if they are independent?  What does it matter if the random 30 year set includes 1871, 1915, 2009, and so on together?  They're supposed to be unrelated.

If the setup of these two websites (and others like it) did have all the permutations of n-year periods rather than chunks that move together through time, what the markets did in the last 5 years truly would not matter or bias the results, no?
Hmm... interesting.  My take on it is that while "past results do not predict future returns" there are correlation between years.  Intuitively this makes sense.  If there is a huge drop in the market it affects virtually all large companies in some shape or form. The following year the companies are influenced by these events.  Customers may not be buying more, but they may take steps to ensure profitability, like laying off employees, squeezing out more productivity, or consolidating their business model.

So... my thought is no.  You need to link the 30 year time periods for it to make any sense.  You can't just pull any collection of 30 years and string them together.    If you do that, by pure chance you'd sometimes put 2008 (-37%), 2002(-37%), and 1974(-42%) all together, which would be catastrophic, and way, way worse than anything that has ever historically happened.  A $10k investment would have dropped to $2,302 in that time frame.

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #24 on: April 10, 2014, 07:52:13 AM »
I apologize if this drifts or I am not understanding a concept but I have a question about firecalc and cfiresim.  If the assumption is that past does not predict the future but does give a reasonable range of what can happen given enough periods of time, why are the 30 year (or whatever number of years) periods ordered?

e.g. Results spit out 113 lines representing the 113 30 year periods but shouldn't we be seeing 113! (factorial) permutations  of 30 year periods if they are independent?  What does it matter if the random 30 year set includes 1871, 1915, 2009, and so on together?  They're supposed to be unrelated.

If the setup of these two websites (and others like it) did have all the permutations of n-year periods rather than chunks that move together through time, what the markets did in the last 5 years truly would not matter or bias the results, no?

Because it is a historical analysis of how your particular circumstances would have performed in any 30 year time frame.

Just salad tossing all the data would be an interesting approach but a Monte Carlo simulation would do better.

mrigney

  • Stubble
  • **
  • Posts: 157
    • Running of the Fools
Re: firecalc and cFIREsim both lie?
« Reply #25 on: April 10, 2014, 08:09:02 AM »
cfiresim and firecalc are a form of a Monte Carlo simulation, fyi. They just happen to sample all available starting dates instead of some population of all available data.

The Vanguard link posted earlier is also a MC sim of the sort described by simonsez. For a 30 year retirement period they randomly select a rate of return for year 1. Then randomly select a rate of return for year 2. Through all 30 years. Then they repeat 5,000 times. That gives a pretty complete range of possibilities:-)


Cpa Cat

  • Handlebar Stache
  • *****
  • Posts: 1598
Re: firecalc and cFIREsim both lie?
« Reply #26 on: April 10, 2014, 08:12:36 AM »
Those last five years will only be in a few of your 30-year projections. I run mine on 70 years (in case I live to be 100+). I get 100% success. At my least successful, I die with under $1m in the bank. At my most successful, I die with 180 MEEEELLION DOLLARS.

If anything, your data only gets better for every 5 years that passes - regardless of how the market does. Because now you have 5 more 30-year ranges to run. You didn't lose five scenarios, you gained 5 scenarios. More data is usually better when trying to run averages.

It doesn't matter if the last five years are outliers - because so is the Great Depression, and it's in there too. Overall, the more data you have, the less impact outliers have.

Cpa Cat

  • Handlebar Stache
  • *****
  • Posts: 1598
Re: firecalc and cFIREsim both lie?
« Reply #27 on: April 10, 2014, 08:24:13 AM »
I would also like to point out that the last five years are essentially irrelevant data today.

It has produced 5 new scenarios in which you have a huge upswing in the last 1-5 years before death. This would be highly unlikely to impact your life in any way.

In order to get a skewed result, you'd have to be running this simulation 30 years from now, at which point there would be 5 scenarios available in which you retire and immediately experience a huge portfolio surge.

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #28 on: April 10, 2014, 08:27:22 AM »
cfiresim and firecalc are a form of a Monte Carlo simulation, fyi. They just happen to sample all available starting dates instead of some population of all available data.

The Vanguard link posted earlier is also a MC sim of the sort described by simonsez. For a 30 year retirement period they randomly select a rate of return for year 1. Then randomly select a rate of return for year 2. Through all 30 years. Then they repeat 5,000 times. That gives a pretty complete range of possibilities:-)

I don't think cfiresim and firecalc are a form of Monte Carlo simulation. If I run the same numbers I will always get the same answer in cfiresim. That is a deterministic system.

A true Monte Carlo simulation will give a different answer each time you put in the same numbers as it will run different "scenarios" or rather just give different variables.

As I said before cfiresim is a historical look. It's quite valuable when you go under the assumption that nothing in the future can be worse than the past (optimism gun), but understand that doesn't mean that rough times can't happen (reality fish smack to the face).

bo_knows

  • Pencil Stache
  • ****
  • Posts: 819
  • Age: 38
  • Location: Fairfax, VA, USA
    • The Crowdsourced FIRE simulator
Re: firecalc and cFIREsim both lie?
« Reply #29 on: April 10, 2014, 08:49:33 AM »
"Lie" is a strong word ;)

I think that these tools show plenty of historical value in their analysis, but should be only used as part of the decision making process.  As Pfau's studies have confirmed (and Sol has theorized), where you retire in the bull/bear cycle matters hugely, but it does not invalidate the historical analysis of cFIREsim or FireCalc IMO.

mrigney

  • Stubble
  • **
  • Posts: 157
    • Running of the Fools
Re: firecalc and cFIREsim both lie?
« Reply #30 on: April 10, 2014, 08:50:49 AM »
@matchweed If you think of every historical year as a member of a population, you could turn firecalc/cfiresim into a MC sim by randomly sampling from those years (say 30, 40, 50 or however many times). Each sampling would provide slightly different initial conditions (as a function of year). However, cfiresim and firecalc always sample every possible initial condition. You're definitely right that a true MC would give a different result each time you ran it. In a traditional MC, you're also typically perturbing initial conditions for each run, starting the sim with a different random number seed, etc. So I guess I'm agreeing that they aren't MC Sims in the traditional sense, but a sort of bastardized form where they have in essence sampled every possible "random" sample. Maybe it'd be better to call it an ensemble "forecast" (even though I railed against calling it a forecast earlier). With that said, as someone pointed out earlier, I'm not sure I buy into Vanguard's MC Sim methodology. The idea of sampling years randomly to get your 30 year ROI seems to assume that year-to-year returns are uncorrelated. Is that actually true? I haven't run the numbers. Might be true. If it's true that year-to-year returns are uncorrelated, it would also mean that a 5 year bull market shouldn't change the validity of a "forecast" from firecalc (since next year's returns would be uncorrelated to the recent bull market).

bo_knows

  • Pencil Stache
  • ****
  • Posts: 819
  • Age: 38
  • Location: Fairfax, VA, USA
    • The Crowdsourced FIRE simulator
Re: firecalc and cFIREsim both lie?
« Reply #31 on: April 10, 2014, 09:02:59 AM »
@matchweed If you think of every historical year as a member of a population, you could turn firecalc/cfiresim into a MC sim by randomly sampling from those years (say 30, 40, 50 or however many times). Each sampling would provide slightly different initial conditions (as a function of year).

Ehhh, you don't want to really do that.  It's important to keep the historical data in series, because the return/inflation/etc is all correlated from what was going on in the world at the time.  Randomizing those data points wouldn't make sense.

FWIW, though, both cFIREsim and FireCalc have actual Monte Carlo options (though, neither is robust)

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #32 on: April 10, 2014, 09:05:48 AM »
@matchweed If you think of every historical year as a member of a population, you could turn firecalc/cfiresim into a MC sim by randomly sampling from those years (say 30, 40, 50 or however many times). Each sampling would provide slightly different initial conditions (as a function of year). However, cfiresim and firecalc always sample every possible initial condition. You're definitely right that a true MC would give a different result each time you ran it. In a traditional MC, you're also typically perturbing initial conditions for each run, starting the sim with a different random number seed, etc. So I guess I'm agreeing that they aren't MC Sims in the traditional sense, but a sort of bastardized form where they have in essence sampled every possible "random" sample. Maybe it'd be better to call it an ensemble "forecast" (even though I railed against calling it a forecast earlier). With that said, as someone pointed out earlier, I'm not sure I buy into Vanguard's MC Sim methodology. The idea of sampling years randomly to get your 30 year ROI seems to assume that year-to-year returns are uncorrelated. Is that actually true? I haven't run the numbers. Might be true. If it's true that year-to-year returns are uncorrelated, it would also mean that a 5 year bull market shouldn't change the validity of a "forecast" from firecalc (since next year's returns would be uncorrelated to the recent bull market).

Unless you feel that you can guess what the market will do then yeah there is no correlation between one year's returns or the next. Now do other variables have any correlation from one year to the next? Dunno. Or do the variables correlate with each other? Well bo_knows... (lame I know).

I also wouldn't call it a forecast. Just call it what it is, an analysis of how a certain situation would perform historically.

mrigney

  • Stubble
  • **
  • Posts: 157
    • Running of the Fools
Re: firecalc and cFIREsim both lie?
« Reply #33 on: April 10, 2014, 09:08:32 AM »
@matchwed Yeah, that was my contention in my original post. To use firecalc as a forecast of some sort is flawed. The "success rate" that it spits out is really just a different way of summarizing historical data. I don't think we're in disagreement really....and I'll concede that it's not a real MC....maybe MC like;-)

bo_knows

  • Pencil Stache
  • ****
  • Posts: 819
  • Age: 38
  • Location: Fairfax, VA, USA
    • The Crowdsourced FIRE simulator
Re: firecalc and cFIREsim both lie?
« Reply #34 on: April 10, 2014, 09:11:59 AM »
@matchwed Yeah, that was my contention in my original post. To use firecalc as a forecast of some sort is flawed. The "success rate" that it spits out is really just a different way of summarizing historical data.

Just FYI, neither of those tools directly state that it's for forecasting the future.  Specifically they state something along the lines of "If the future is no worse than the worst years of the past, then your portfolio will survive".  I think people often warp the meaning of the charts to think that it's a future forecast.

PeteD01

  • Bristles
  • ***
  • Posts: 415
Re: firecalc and cFIREsim both lie?
« Reply #35 on: April 10, 2014, 09:38:05 AM »
Interesting topic.

I think that an individual who is close to FIRE and has enough education and experience under their belt should be able to pick up on the risks inherent in FIRE. It is already known that a market drop at the beginning of FIRE is a risk which cuts your success rate. And of course it is a (I don't want to use the word flaw or a lie, maybe downside?) downside to historical data usage that more successful years inflate your success rate while more low years inflate your failure rate.

I guess we just have to be those educated and experienced people. :)

As usual we can't take all the advice from the boards on blind faith. These are things we have to grok.

Exactly.

The recent market returns have put me ahead of schedule and have possibly increased the sequence of return risk at the same time. And that is what the problem is about: The influence of market performance immediately preceding retirement on sequence of returns risk

I'm very much into the life cycle approach and I'm dealing with the issue in the following way:

Set the future earnings potential at the day of RE to zero - that's retirement alright.

Match liabilities in two ways:

1. reduce cash flow needs by paying off mortgage/selling other liabilities

2. match essential cash flow needs with guaranteed/rental income. I sold stock funds and bought deferred fixed annuities (no inflation rider - it is not necessary when dealing with early sequence of returns risk and initial payout is much higher) and have a small rental income.

Check size of the market investments with respect to the ability to support desirable level of discretionary spending.

Add it all up and match the shortfall by realizing future earning potential + investment returns and one gets a date of FIRE.

Result in my case: 12 months more time to FIRE as of last summer, independent of market returns (I have enough already in the upside portfolio). A 4% withdrawal rate and an assumption of constant 6-7% market returns would have meant FIRE last summer but, because the numbers were so close, any significant downturn would have been a problem and I would have had to work for another year anyway to be on the safe side.

I my particular case, the decreasing present value of my non inflation adjusted annuities will lead over time to an increasing equity allocation and the additional incomes (pension, SS) kicking in over the next 18 years will make the market risk acceptable.
I have traded some upside potential for a firm date of FI and have set up for a rising equity allocation throughout retirement.
My future earnings potential is not really zero at the date of FIRE but belongs now strictly to the upside as does the market performance.

My plan is a little more complicated in the details but those vary anyways between households. I have access to attractive TIAA CREF annuities but for others bond/annuity ladders may be more attractive. In any case, this is just an example from someone close to FIRE.

Peter

jordanread

  • Guest
Re: firecalc and cFIREsim both lie?
« Reply #36 on: April 10, 2014, 09:49:03 AM »
Don't mind me, just commenting so I can catch up on it. I'll be doing some work with both of those soon, so I'll chime in once I have more data.

NumberCruncher

  • Pencil Stache
  • ****
  • Posts: 612
Re: firecalc and cFIREsim both lie?
« Reply #37 on: April 10, 2014, 09:57:04 AM »
I apologize if this drifts or I am not understanding a concept but I have a question about firecalc and cfiresim.  If the assumption is that past does not predict the future but does give a reasonable range of what can happen given enough periods of time, why are the 30 year (or whatever number of years) periods ordered?

e.g. Results spit out 113 lines representing the 113 30 year periods but shouldn't we be seeing 113! (factorial) permutations  of 30 year periods if they are independent?  What does it matter if the random 30 year set includes 1871, 1915, 2009, and so on together?  They're supposed to be unrelated.

If the setup of these two websites (and others like it) did have all the permutations of n-year periods rather than chunks that move together through time, what the markets did in the last 5 years truly would not matter or bias the results, no?

This is basically what the vanguard link mentioned above does, except it just randomly picks the next year (random 1-year periods only as opposed to random n-year chunks): https://retirementplans.vanguard.com/VGApp/pe/pubeducation/calculators/RetirementNestEggCalc.jsf 

If you have a high probability of success with that calculator, there's just about nothing that can slow you down (unless the markets experienced all the historical financial crises one after the other). These results seem much more conservative than cfiresim.

sirdoug007

  • Pencil Stache
  • ****
  • Posts: 587
  • Age: 39
  • Location: Houston, TX
Re: firecalc and cFIREsim both lie?
« Reply #38 on: April 10, 2014, 10:17:14 AM »
The times a 4% withdrawal rate has failed according to cFIREsim and the Shiller PE on Jan 1 of that year:
1906: 20.1
1965: 23.27
1966: 24.06
1967: 20.43
1968: 21.51
1969:21.19
1973: 18.71

The average PE10 all-time is 16.52, so these are all at least 10% higher than the average.  The 4% rule has failed in 7/39 cases where the Shiller PE was above the average, and 6/15 times that the starting PE10 was above 20.

The problem with this thinking is that it would lead you to never retire.  The last time the Shiller PE was less than 20 was 2009.  Before that it was 1992. 

So between 1992 and 2009, 17 years, you would have still been waiting for the "Big One".  Even in the aftermath of the tech bubble, the Shiller PE dropped only to about 21.

I think the MMM approach of covering your expenses with a very reasonable but not 100% safe withdrawal rate of 4% and rolling with the punches from there is the way to go.

If the market does well, good stuff.  If it doesn't, maybe go find a way to make some side money to reduce your withdrawal rate during those downturns.

You can spend a lot of time trying to predict the future.  However, the one thing that is guaranteed is that your projection will be wrong. :)

bo_knows

  • Pencil Stache
  • ****
  • Posts: 819
  • Age: 38
  • Location: Fairfax, VA, USA
    • The Crowdsourced FIRE simulator
Re: firecalc and cFIREsim both lie?
« Reply #39 on: April 10, 2014, 10:21:54 AM »
I think the MMM approach of covering your expenses with a very reasonable but not 100% safe withdrawal rate of 4% and rolling with the punches from there is the way to go.

+1 to that.  This is why I like modeling my expenses with a "floor" value that represents barebone necessities, and a "ceiling" value that represents a comfortable retirement.  If cFIREsim tells me that a model like that passes, I feel much better than a strict 4%.

arebelspy

  • Administrator
  • Senior Mustachian
  • *****
  • Posts: 28025
  • Age: -999
  • Location: Seattle, WA
Re: firecalc and cFIREsim both lie?
« Reply #40 on: April 10, 2014, 10:35:02 AM »
So between 1992 and 2009, 17 years, you would have still been waiting for the "Big One".

Not necessarily.  He's saying a 4% SWR may fail at 20+ PE10, but presumably while keeping working waiting for a crash you'd accumulate enough to get down to a 3.5% or 3% SWR and be able to FIRE despite not having a crash or not hitting a PE10 of < 20.

It's not "work until crash" if you can get your SWR below that 4% SWR.
We are two former teachers who accumulated a bunch of real estate, retired at 29, spent some time traveling the world full time and are now settled with two kids.
If you want to know more about us, or how we did that, or see lots of pictures, this Business Insider profile tells our story pretty well.
We (rarely) blog at AdventuringAlong.com. Check out our Now page to see what we're up to currently.

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #41 on: April 10, 2014, 11:17:46 AM »

This is the exact same problem another forum member mentioned here recently, about how everyone who is invested in the market is much more likely to hit their "target number" in the middle of a market surge than during a downturn, which virtually guarantees that we're all going to think we're retirement eligible based on overvalued assets.


The problem is t

But without considering the correlating trend between inflation a

arebelspy

  • Administrator
  • Senior Mustachian
  • *****
  • Posts: 28025
  • Age: -999
  • Location: Seattle, WA
Re: firecalc and cFIREsim both lie?
« Reply #42 on: April 10, 2014, 11:21:47 AM »

This is the exact same problem another forum member mentioned here recently, about how everyone who is invested in the market is much more likely to hit their "target number" in the middle of a market surge than during a downturn, which virtually guarantees that we're all going to think we're retirement eligible based on overvalued assets.


The problem is t

But without considering the correlating trend between inflation a

Another site bug?  I'm not sure your posts ar
We are two former teachers who accumulated a bunch of real estate, retired at 29, spent some time traveling the world full time and are now settled with two kids.
If you want to know more about us, or how we did that, or see lots of pictures, this Business Insider profile tells our story pretty well.
We (rarely) blog at AdventuringAlong.com. Check out our Now page to see what we're up to currently.

PeteD01

  • Bristles
  • ***
  • Posts: 415
Re: firecalc and cFIREsim both lie?
« Reply #43 on: April 10, 2014, 11:47:06 AM »

This is the exact same problem another forum member mentioned here recently, about how everyone who is invested in the market is much more likely to hit their "target number" in the middle of a market surge than during a downturn, which virtually guarantees that we're all going to think we're retirement eligible based on overvalued assets.


The problem is the "target number". When looking at retiring at a specific (age driven) date of retirement, the sequence of return risk is determined by the date of birth - more or less. When the date of retirement is determined by a "target number" (market driven - more or less) the risk of retiring towards the end of a bull market with consequently increased sequence of return risk is increased. The probability of success based on historical back testing the simulators spit out applies to the population of age driven retirees. It is therefore incorrect to use the result for a population of market driven retirees who are more likely to retire after significant market advances. Their probability of success is lower but cannot be quantified because on the day of the simulation it is not at all clear if the retiree indeed will have retired towards the end of a bull market.
Now, one could assume that the end of the bull market is near and then back test using only similar periods and I believe Wade Pfau is one who has done something in this direction. The projections should look much more mediocre then.
I would not interpret the overly optimistic projections of these simulators as a design flaw or bias, but rather acknowledge that they were not designed to be used in the same way by market driven as by age driven retirees.

Peter

PeteD01

  • Bristles
  • ***
  • Posts: 415
Re: firecalc and cFIREsim both lie?
« Reply #44 on: April 10, 2014, 11:50:14 AM »

This is the exact same problem another forum member mentioned here recently, about how everyone who is invested in the market is much more likely to hit their "target number" in the middle of a market surge than during a downturn, which virtually guarantees that we're all going to think we're retirement eligible based on overvalued assets.


The problem is t

But without considering the correlating trend between inflation a

Another site bug?  I'm not sure your posts ar

I don't know what happened and I messed it up even more by removing the rogue post...

Exflyboy

  • Walrus Stache
  • *******
  • Posts: 6401
  • Age: 58
  • Location: Corvallis, Oregon
  • Expat Brit living in the New World..:)
Re: firecalc and cFIREsim both lie?
« Reply #45 on: April 10, 2014, 11:55:06 AM »
Yes the problem of a dramatic pullback right after you retire has not been lost on me.

My approach is to have about 2 to 3 years of expenses in cash, plus about half of our expenses are covered by rental income.

I'm hoping if we have a big pullback this year I will be able to ride through without drawing on any of the investments.

Frank

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #46 on: April 10, 2014, 11:56:09 AM »
Hmm, I'm not sure if an age driven retirement even exists. It's rather arbitrary as it doesn't necessarily say anything about how successful you will be in retirement regardless of market return. Your assets, cash flow, market performance, inflation...etc. all factor in much more than how old you are.

simonsez

  • Pencil Stache
  • ****
  • Posts: 796
  • Age: 33
Re: firecalc and cFIREsim both lie?
« Reply #47 on: April 10, 2014, 11:59:04 AM »
Hmm, I'm not sure if an age driven retirement even exists. It's rather arbitrary as it doesn't necessarily say anything about how successful you will be in retirement regardless of market return. Your assets, cash flow, market performance, inflation...etc. all factor in much more than how old you are.
Sounds like you are avoiding the golden handcuffs quite well.

PeteD01

  • Bristles
  • ***
  • Posts: 415
Re: firecalc and cFIREsim both lie?
« Reply #48 on: April 10, 2014, 12:03:20 PM »
Hmm, I'm not sure if an age driven retirement even exists. It's rather arbitrary as it doesn't necessarily say anything about how successful you will be in retirement regardless of market return. Your assets, cash flow, market performance, inflation...etc. all factor in much more than how old you are.

I am not necessarily implicating that an age driven retirement exists in reality or not.
The point is that the models used in the simulators assume retirement dates independent of preceding market returns.
The real world obviously contains everything from the retire at sixty five to somewhat market driven to completely market driven.

Peter

matchewed

  • Magnum Stache
  • ******
  • Posts: 4334
  • Location: CT
Re: firecalc and cFIREsim both lie?
« Reply #49 on: April 10, 2014, 12:32:28 PM »
Hmm, I'm not sure if an age driven retirement even exists. It's rather arbitrary as it doesn't necessarily say anything about how successful you will be in retirement regardless of market return. Your assets, cash flow, market performance, inflation...etc. all factor in much more than how old you are.
Sounds like you are avoiding the golden handcuffs quite well.

Hah, considering tomorrow is my last day on the job before school starts in the fall I guess I am. :)

That being said it's not necessarily the age that is the factor in those incentives but the money no?

Hmm, I'm not sure if an age driven retirement even exists. It's rather arbitrary as it doesn't necessarily say anything about how successful you will be in retirement regardless of market return. Your assets, cash flow, market performance, inflation...etc. all factor in much more than how old you are.

I am not necessarily implicating that an age driven retirement exists in reality or not.
The point is that the models used in the simulators assume retirement dates independent of preceding market returns.
The real world obviously contains everything from the retire at sixty five to somewhat market driven to completely market driven.

Peter

Well what I mean is shouldn't all retirement attempts be market driven (I'll be honest I'm not sure I even like the term given my want for nuance and understanding risks and pitfalls as people get closer to retirement)?