Request for critique: blackjack simulator methodology, validation, and CVData gaps

I’d like to ask for some informed criticism on a blackjack simulator / analysis project I’ve been building.

The site is here:

https://ether-ore.github.io/BJW/

This is not a sales post, and I’m not claiming CVData/CVCX parity. In fact, one reason I’m posting is that I do not yet have that parity, and I’d rather be candid about that.

I also want to be clear that I’m not trying to outdo CVData/CVCX. I think Blackjack Wonk may be aimed in a somewhat different direction, though I’m still trying to define that clearly myself: less toward replacing existing software, and more toward exploring questions and experiments that I want to be able to set up in my own way.

The public site lays out the simulator structure, methodology, validation work, references, and unresolved gaps.

If any experienced players or experienced users of blackjack software are willing to look it over, I’d especially value thoughts on:

- where you would investigate first if you were trying to reconcile the remaining CVData/CVCX differences
- whether the methodology/validation work looks sound or still has obvious holes
- whether there are blackjack questions current software does not handle especially well, where something like this might be useful

I’d much rather hear blunt criticism than praise.

Thanks to anyone willing to take a look.

Can you tell me, please, where you would find this formula in chapter 10??

BJW implements the Schlesinger Chapter 10 variance formula directly:

Var(session) = E[b²] · Var(w) + Var(b) · E[w]² + 2 · Cov(b,w) · E[b] · E[w]

Don

Don, Having only 5 posts on general pages, and NONE on Green Chip,  the question is why a novice person as such would not devote his memorized energy to proven indices as yours, Wong, or Revere?

Your gracious time/energy, IMHO,  goes for naught to his foolishness!

Anyone beg to differ?

Qqq

Don, you’re right to call me out here. I can’t currently cite that exact equation to BJA3, so I should not have labeled it a Chapter 10 formula from Schlesinger. At best, “Chapter 10” was shorthand for the broader variance/DI/SCORE framework, but the specific equation in our docs is our own implementation choice, not a verified direct citation from your book.

I’m scrubbing that language from the documentation and related materials wherever it appears so we don’t repeat the mistake or imply an attribution we haven’t actually sourced. If I later find an exact citation, I’ll use the citation; if not, I’ll describe it only as our own formulation. Thanks for calling it out and my sincere apologies.

IT APPEARS YOU JUST DON'T GET IT!!

In addition to Expected Value (EV) results please highlight Variances and Standard Deviations. Outcomes of a casino visit and even many casino visits are very random and averages are not typical of real world results. EV per hand or EV per hour are also extremely random. 

Thank you,

Phil

I agree with Phil. These statistics numbers, EV and Standard Deviation, depend strongly on how they are calculated. In a single-deck blackjack game, there are 13x13=169 ways of player’s starting hands; therefore, there are 169x13=2197 ways of player-dealer combinations. The unit of statistics should be greater than 2197 hands, to say the least. 

Phil and aceside — both are fair critiques.

Phil: I agree EV alone is not enough. A real casino trip is one noisy draw from a wide distribution, so expectation by itself tells you very little about what an actual session will look like. BJW does track the variance side too — SD, DI, SCORE, and N₀ — precisely because EV alone is not an adequate description of real-world blackjack results.

aceside: I agree with the substance of your point as well. EV and SD absolutely depend on how they are defined and calculated. I would not use 2,197 hands as a real convergence floor, but I do agree with the principle behind it: very small samples can produce numbers that look precise without being trustworthy. In practice we do not treat tiny runs as meaningful at all; serious comparison work has generally been pushed into the 100M-round range, and canonical validation is usually reserved for much larger runs of 1 billion rounds or more.

For the standard definitions of DI, SCORE, and N₀, QFIT’s public CVCX definitions page states them in essentially the same form.

There are very few people who will ever play even a couple of hundred thousand hands of blackjack in their lifetime.

Phil

I recently tried a new blackjack simulator,  bjtheorem 'dot' com, from an international team and found it interesting. It seems to me they use 16 million hands in each simulation. Also, they break each simulation into equal segments of 2000 hands each and then produce EV and SD statistics on each simulation. The result is that these EV and SD numbers also depend on the number of players at the table. 

To hopefully clarify more, the details about what is contained in the manifest and results files and how to read them can be found here:

https://ether-ore.github.io/BJW/read-manifest-json/

I’m learning. In your manifest, "burn_cards": 1, do you mean the first card of each card shoe is exposed to the players? 
 

Why do you even need this? You just start dealing without burning to simplify everything. 

No. It means one card is burned unseen after the shuffle, not exposed. The burn card does not affect the count. We include it because that is part of the real dealing procedure, and it does slightly change the remaining shoe composition even if players never see the card.

In that case, you can just set "burn_cards"=0 all the time. I’ve tried a few blackjack simulators, including Phil’s Blackjack-hack, but none of them includes a burn card option. Your consideration of variance calculation is interesting. Why not focus on this side of simulation? 

You could do that, but I would ask why? What actual blackjack table doesn't burn a card? Not saying it's impossible but I haven't seen one. 

As to burn cards, it's in CVData as well.


 

No disagreement there. I run a billion rounds not because I think anyone will ever play that many, but to reduce the deviation to a level where the numbers are reliable. 

The purpose of burning a card is to deter card marking in casinos.  In blackjack simulation, it’s redundant. Furthermore, it introduces complications and more questions, just like the one I asked you above. 

Ok

Another benefit of no burning card is toward better comparison. Your EV and SD numbers can be easily compared to Phil’s with a fixed 75% penetration point. If you have a burn card, your penetration would be 74.679% ?

blackjack-hack burns one card at the beginning of each shoe, not visible to player

https://blackjack-hack.com

Phil