vpFREE2 Forums

Risk of Ruin before Royal simulation

Steve Jacobs wrote (snip):

he following table compares the three figures you've provided with
the corresponding values from my calculations (which I still claim to
be exact):

bank Paymar Jacobs Paymar/Jacobs
-----------------------------------------------
2000 13.1 14.99 0.874
1000 36.2 38.72 0.935
984 50.0 39.31 1.272
-----------------------------------------------

Steve, you've taken these figures from two different posts. The 984 bankroll for 50% RORBR came from a very early post. That is clearly wrong. My latest figure from a 10,000-session simulation is 740 units, which I believe is quite close to your number.

Also, my latest 10,000-session simulation for a 2000-unit bankroll is 13.7%. It seems we're getting closer.

If my numbers a truly correct as I believe they are, then your formula
is more accurate for the 1000 unit bankroll than it is for values above
or below 1000 units. This might also suggest that your formula isn't
really computing RORBR, but something quite different. Perhaps we
should examine a few more data points?

The number you've posted for RORBR(1) (i.e., Risk of Ruin Before Royal starting with a bankroll of one betting unit) shows nine significant figures. Perhaps that is not sufficient for an accurate value for RORBR(N) when N is large.

For example, you give RORBR(1) = 0.999051654. I find that 0.999051654^50 = .9536 or 95.36%, while my simulation for a 50-unit bankroll gives 95.4%, a perfect match within the tolerance of the simulation.

I agree that we need to continue examining this and resolve the differences.

Dan

···

--
Dan Paymar
Author of best selling book, "Video Poker - Optimum Play"
Editor/Publisher of VP newsletter "Video Poker Times"
Developer of VP analysis/trainer software "Optimum Video Poker"
Visit my web site at www.OptimumPlay.com

"Chance favors the prepared mind." -- Louis Pasteur

Let's take a look at the "Sorokin" RoR formula. (Please don't get hung up on the who first
applied the formula to VP. It just doesn't matter here) In a future email I will take a good
look at the RoRBR formula.

How much confidence to we have that the RoR formula is correct? Do we really know what
RoR represents? To begin to answer these questions, I am going to consider an incredibly
simplified version of VP. Though it is simplified , the VP game I will use captures all the
"elements" (I like to call it the "physics") necessary to understand RoR.

----------------------------------- The game

···

----------------------------------------

So here is the game: Instead of multiple outcomes, there are only 2 as follows:

    Outcome 0: No return (bet is lost) with a Probability of P0
    Outcome 1: A return of w units with a Probability of P1, W>0

With Po+ P1 = 1 and P0>0 and P1>0

Only 1 unit can bet at a time. The EV of the game is EV = 0 * P0 + W*P1 - 1 (I subtract
the 1 so that EV is either positive, negative or zero)

Please don't balk at the use of such a simple game. Instead, be patient, and see where it
leads us. We will add some more complexity to the game later.

For now, I am going to assume (without loss of generality) that we have a 1 unit bankroll
and that we only bet 1 unit at a time (I already said that).

----------------------------------- RoR Formula
----------------------------------------

If we apply the RoR formula to this simple game, we get

     RoR = x = P0 + P1 * x ^w

For w=1, this takes on a particularly simple form

   RoR(w=1) = P0/ (1-P1) = 1 since 1-P0 = 1

For w =2, the solution is also readily available, but a little more complex

    RoR(w=2) = [1 +/- sqrt( 1 - 4*P1*P0)] / 2*P1

Noting that PO = 1 - P1, we get,

   RoR(w=2) = [1+/- (2P1-1) ]/2P1

Which produces 2 values,

   RoR(w==2) =1 or RoR(w=2) = (1-P1)/ P1 = P0/P1

Let's limit ourselves to these simple cases, where w=1 and w=2.

So what have we learned so far?

For w=1, RoR =1. That is, for this simple game, ruin is inevitable. That makes sense. The
game is NEGATIVE, so RoR =1. Good

For w=2. RoR can take on 2 values, RoR=1 or RoR = P0/P1. Since w=2, the game is
positive, so we expected 2 answers, and one that said RoR <1 and we got it

----------------------------------- Other approach
----------------------------------------

Let's look at the calculation of RoR from first principles.
After any number of hands there are only 2 possible "states": Either we are "still playing"
or we are "ruined"
So,

     P(ruined,n) + P(still playing,n) = 1

where

    P(ruined, n) is probability that we are ruined after hand n
    P(still playing, n) is probability that we are still playing after hand n

This is a very important result that bears repeating: There are only 2 states to this "game",
and since we must always be in one of the state, the sum of their probabilities must equal
1.

Now, let look at the case when w =1.
At the first play, we can either "Ruin" with a probability of P0 or be still playing, with a
Probability of P1,

   P(ruin,1) = P0
   P(still playing) = P1

Adding up the two probabilities yields,

    P(ruin,1) + P(still playing) = P0 +P1 =1

as it should.

After the first hand, if we are still playing, we again have the possibility of being ruined or
not. Interestingly, the contributions from each play to the probabliity of Ruin are additive,
while those for "still playing" are multiplicative. (Please don't take my word for this, work
it out for yourself). So,

   P(ruin,2) = P(ruin,1)+ P0 * P1 = P0 + P0*P1

   P(still playing) = P(still playing, 1) *P1 = P1^2

By extension, after the n-th hand, we have

   P(ruin,n) = P0 + P0*P1 + P0*P1^2 + .... + P0*P1^(n-1) = P0 * [ 1 + P1 + P1^2 + .... +
P1^(n-1)]

   P(still playing) = P1+ P1^2 + ....+ P1^n = P1* [ + P1 + P1^2 + .... + P1^(n-1)]

Both of these geometrical series have well known formulas, so,

   P(ruin,n) = P0 * ( 1 -P1^n)/ (1-P1)
   P(still playing,n) = P1 * ( 1 -P1^n)/ (1-P1)

We can check that the sum of the two probabilities is equal to 1,

   P(ruin,n) +P(still playing,n) =? 1
   P0 * ( 1 -P1^n)/ (1-P1) + P1 * ( 1 -P1^n)/ (1-P1) = (P0 + P1) * *( 1 -P1^n)/ (1-P1) = 1

as it should.

Let's simplify R(ruin,n),

   P(ruin,n) = P0 * ( 1 -P1^n)/ (1-P1) = 1-P1^n since PO = 1 -P1

So we have arrived at our answer,

  P(ruin,n) = 1 -P1^n

Earlier we found that, using a different approach,

   RoR(w=1) =1

By noting that, in the limit of n->infinity,

  P1^n -> 0 since P1< 1,

so,

   P(ruin,n->infinity) = 1

So, at this point we can conclude, RoR indeed gives the P(ruin,n) for infinite n (for w=1, a
negative game).

Though this is a trivial result, it sure is reassuring.

END OF PART 1. In the next part, I hope to tackle P(ruin, w=2, n)

Let's take a look at the "Sorokin" RoR formula. (Please don't get

hung up on the who first

How much confidence to we have that the RoR formula is correct?

Can't speak for "we", but I'm 100% confident it's correct.

Do we really know what
RoR represents?

Well, I'd hope anyone who gambles understands the concept of gambler's
ruin.

For w=2. RoR can take on 2 values, RoR=1 or RoR = P0/P1. Since

w=2, the game is

positive, so we expected 2 answers, and one that said RoR <1 and we

got it

Not quite. There are indeed two roots. A probability has to satisfy
0<=p<=1. So if P0>P1, or equivalently P0>1/2 the relevant root is 1
and the risk of ruin is 1. If P0<1/2 the relevant root is P0/P1 and
the risk of ruin is less than 1.

You know your time and that of anyone who reads your posts might be
better spent proving to yourself that the "Sorokin" formula is correct
and that a straightforward extension of the logic applies to the risk
of ruin before a jackpot problem. To derive the result you only need
two results from chapter 1 of any probability textbook, namely that
the probability of the union of two disjoint events A and B is

P(A OR B) = P(A)+P(B)

and the probability of two independent events is

P(A AND B) = P(A)P(B)

That's all you need. The rest is just partitioning the event space
into the appropriate collection of disjoint events and independent events.

Mike

···

--- In vpFREE@yahoogroups.com, "cdfsrule" <groups.yahoo@v...> wrote:

Steve Jacobs wrote (snip):
>he following table compares the three figures you've provided with
>the corresponding values from my calculations (which I still claim to
>be exact):
>
>bank Paymar Jacobs Paymar/Jacobs
>-----------------------------------------------
>2000 13.1 14.99 0.874
>1000 36.2 38.72 0.935
> 984 50.0 39.31 1.272
>-----------------------------------------------

Steve, you've taken these figures from two different posts. The 984
bankroll for 50% RORBR came from a very early post. That is clearly
wrong. My latest figure from a 10,000-session simulation is 740
units, which I believe is quite close to your number.

You've missed the point. The figures in the table compare your
NON-simulation numbers with the numbers from my formula. The
fact that 984 came from an earlier post should only matter if you
have _changed_ your RORBR method in mid-discussion.

Was your earlier figure of 984 a mistake, or have you changed
your RORBR method? In order to be consistent with your 2000-unit
and 1000-unit numbers, the 50/50 bankroll would need to be about
682 units.

Also, my latest 10,000-session simulation for a 2000-unit bankroll is
13.7%. It seems we're getting closer.

>If my numbers a truly correct as I believe they are, then your formula
>is more accurate for the 1000 unit bankroll than it is for values above
>or below 1000 units. This might also suggest that your formula isn't
>really computing RORBR, but something quite different. Perhaps we
>should examine a few more data points?

The number you've posted for RORBR(1) (i.e., Risk of Ruin Before
Royal starting with a bankroll of one betting unit) shows nine
significant figures. Perhaps that is not sufficient for an accurate
value for RORBR(N) when N is large.

For example, you give RORBR(1) = 0.999051654. I find that
0.999051654^50 = .9536 or 95.36%, while my simulation for a 50-unit
bankroll gives 95.4%, a perfect match within the tolerance of the
simulation.

Let's test that theory by changing the ninth digit using a 5000 unit
bankroll:

0.999051653 ^ 5000 = 0.0087038934747
0.999051654 ^ 5000 = 0.0087039370356
0.999051655 ^ 5000 = 0.0087039805967

For bankrolls less than 5000 units, 9 figures for RORBR(1) gives
probability of failure that are accurage to at least 7 digits. Clearly,
nine digits is plenty for any simulation comparisons that we're
likely to make.

···

On Friday 11 November 2005 10:38 am, Dan Paymar wrote:

Mike,

Consider this: In the RoRBR problem, there are 3 possible states, not 2 as in RoR. Two of
them are "absorbing" (they are called absobing becuase once you reach them you can't
leave)

    State (1): Hit RF, done playing (absorbing)
    State (2): Ruined, done playing (absorbing)
    State (3): Still playing

And thus:

   P(Hit RF,n) + P(Ruined,n) + P(still playing, n) =1 for all n

I'd hope we can all agree on this much.

I also claim that RoRBR = P(RF,n) for n-> infinity

Also note that,

For a Negative EV game, P(still playing, n) = 0 for n-> infinity
For a Positive EV game, P(still playing, n) >0 for n-> infinity

Steve says that one can use the "sorokin" formula to compute RoRBR by simply NOT
including the contribution for the RF.

He also says that.

   RoRBR = 1 -F

Where F is probability of failure (not hitting the RF before Ruin).

In other words,

RoRBR = 1 - P(Hit RF) = 1 - P(Still Playing) - P(Ruin)

It might be obvious to you at this point that this is not what the RoRBR formula actually
computes. If not read on.

With a little algebra, one can evaluate Steve's formula for RoRBR and show that the above
relationship does not hold for some simplified VP games (or any positive EV game, for
that matter). One can do the evaluation (for the simplified games) without a single
approximation. It can be done "exactly."

Why doesn't the relationship hold? Becuase a "sorokin"-like RoR formula accounts for only
1 absorbing state. But, in the RoRBR problem, there are 2 absorbing states as there are
two ways that the game ends: hit a royal flush or ruin. So, until somone comes up with a
Sorokin-like formula with 2 absorbing states, I stand by my earlier earlier comments:
Either Steve's approach is flawed, or the problem is ill-posed.

That said, for many cases, Steve's approach does give results that are certainly close
enough. I know that. I am not arguing that his "numbers" are not close enough, I am
arguing that the method is not "exact". Yes, I hate it when people claim that something is
"exact" without at least trying to demonstrate it.

BTW, If you are so sure of statistical abilites that you feel that you must lecture me, please
go ahead an come up with a sorokin-like formula for 2 absorbing states (3 states total).
Until you do, I am going to assume that it differs from the formula for 1 absorbing state.

And yes, I'd really like to see that result. So would many other people on this list. With it,
someone could answer the question
"what is the probability I win X before losing Y" without having to resort to simulation or
PDF convolution. I'm sure there's a way, I just don't know it. Do you?

Are you still absolutely certain that the extension from RoR to RoRBR is straightforward?

If so, you ought to be able to derive Steve's RoRBR-formula and prove that Steve's RoBR =
P(RF, done playing) even for a positive EV game. You can work with the postive game of
your choice.

···

--- In vpFREE@yahoogroups.com, "Michael Peck" <mpeck1@i...> wrote:

You know your time and that of anyone who reads your posts might be
better spent proving to yourself that the "Sorokin" formula is correct
and that a straightforward extension of the logic applies to the risk
of ruin before a jackpot problem.

Opps..... I realize I really screwed up my last post and made my argument needlessly
complicated. Sorry. Here is the correct, consice version:

Sorokin says:

  RoR = 1 - P(Still playing)

Steve Says:

  RoRBR = 1 - F ,

where F is the probability of failure. Failure means either the player is still playing or the
player has been ruined, so

RoRBR = 1 - P(still playing) - P(ruin)

and so, using ,
P(still playing) + P(ruin) + P(RF) =1,

I assume this is what Steve is aiming to compute:

RoRBR = 1 - P(RF)

But a Sorokin tyoe RoR computation only computes 1-P(still playing). There are potentially
3 unknowns in this problem, not 2. I can show this with some algebra. Anyone out there
able to derive the Sorokin RoR or Steve's RoRBR?

Opps..... I realize I really screwed up my last post and made my argument
needlessly complicated. Sorry. Here is the correct, consice version:

Sorokin says:

  RoR = 1 - P(Still playing)

Steve Says:

  RoRBR = 1 - F ,

where F is the probability of failure. Failure means either the player is
still playing or the player has been ruined, so

No! RoRBR = F

Probability of failure is probability that you go bust without getting a
royal. For this equation, "still playing" isn't part of the final result.
In terms of the equivalent Markov chain, the two absorbing states
absorb a fraction of states that quickly approaches all cases.

The pure RoR case is different, since it only has one absorbing
state (unless you consider a 2nd absorbing state to exist at infinity).

RoRBR = 1 - P(still playing) - P(ruin)

and so, using ,
P(still playing) + P(ruin) + P(RF) =1,

I assume this is what Steve is aiming to compute:

RoRBR = 1 - P(RF)

But a Sorokin tyoe RoR computation only computes 1-P(still playing). There
are potentially 3 unknowns in this problem, not 2. I can show this with
some algebra. Anyone out there able to derive the Sorokin RoR or Steve's
RoRBR?

There aren't three unknowns, there is only one, the value that I call F,
which is the probability of going broke before hitting a royal.

For 9/6 JoB the non-royal payoffs are (1, 2, 3, 4, 6, 9, 25, 50) and the
RoRBR equation with unknow F is:

F = p(lose) + p(1)*F + p(2)*F^2 + p(3)*F^3 + p(4)*F^4 + p(6)*F^6
   + p(9)*F^9 + p(25)*F^25 + p(50)*F^50

All of the probabilities are know to very high accuracy. Any number of
people here can produce them on their own, or know where to find them.

The equation above can be understood as follows: The probability that
you will eventually fail, when starting with a single coin, is based on the
probabilities involved in the very first hand played. If your first outcome
is to lose, then your bankroll is gone with probability p(lose). If the
first outcome means that the machine returns N units to you, then you
can treat this result as if you are starting N games in parallel, each with
a single unit bankroll. You fail if and only if all N games result in
failure. So, the exact probability of failure after a payoff of N units
must be exactly F^N, where F is the same RoRBR(1) value that is our
only unknown. The equation accounts for all payoffs _except_ the
royal, because hitting the royal means you _didn't_ fail, so the
probability of hitting a royal cannot factor directly into the equation
for failing to hit the royal.

Please think this through carefully. It is a correct mathematical
formulation, and it is indeed exact.

···

On Saturday 12 November 2005 12:12 am, cdfsrule wrote:

Maybe I'm missing something but it seems pretty straightforward.

The risk of losing n bets is the same as losing one bet n times:

R(n)=R(1)^n

The risk of losing one bet is the sum of the (probability of losing
one bet) plus the (probability of winning one bet times the
probability of losing one bet) plus the (probability of winning two
bets times the probability of losing two bets) ...

R(1)=SUM[prob x R(1)^Win] for each possible outcome

Using this formula, R(1) can be solved for iteratively using a
spreadsheet.

···

--- In vpFREE@yahoogroups.com, "cdfsrule" <groups.yahoo@v...> wrote:

Anyone out there
able to derive the Sorokin RoR or Steve's RoRBR?