Following up on the (now locked) Neural Net Lottery Picker thread — the backtest problemPrev TopicNext Topic

• NumberCruncher2

  

  us
  Member #261,885
  June 21, 2026
  16 Posts
  Online
  Jun 23, 2026, 1:00 pm

  ▲▼

  Was reading back through the old Neural Net Lottery Picker thread and it got locked before the most interesting part really got going. The MCMC/Markov discussion in there was the best thread I've found on here, so starting a fresh one to keep it alive.

  The point that stuck with me: someone noted they found no real advantage running Markov on draw history because the independence of each draw keeps failing to reject "it's just a discrete uniform distribution." That matches what I keep hitting - the models will happily find structure, but it doesn't survive once you hold back a test set. Classic overfit.

  Which lands on the thing I think actually matters more than the model: the backtest. Before trusting any of this, I've started running it backwards over full draw history to see how often the real winning line would've actually survived the filter or ranked well under the model. More often than not, the fancy approach does no better than flat random - which is its own kind of useful to know.

  So for the people in here actually building this stuff:

  • When you test an ML or MCMC approach, do you hold back an out-of-sample set, or judge it on the full history?
  • Has anyone gotten any method to beat uniform-random out-of-sample, even slightly?
  • The IQR hot/cold/neutral classification idea someone raised — anyone actually backtested whether the "hot" bucket outperforms?

  Not looking to relitigate "it's all random" — I know where that lands. More interested in how you separate a method that genuinely does something from one that just looks impressive on the data it was built on.
• BobP

  

  Dump Water Florida
  United States
  Member #380
  June 5, 2002
  3,680 Posts
  Offline
  Yesterday, 1:51 am

  ▲▼

  What I do know about testing lottery software, is you grind your way through the draw history seeing if there is any path to a winning combination(s) over a budget able length of play.  If the lottery eludes every prediction, move on, try another prediction engine.  One upcoming problem is name brand AI has been implanted with the notion selecting winners is impossible and it should only look to sites that are "expert" in the field, ie: random.  When the day comes it can write decent software, it will likely refuse to write lottery software.
  
  BobP
  
  It is impossible to outwit a cat. 
  
   
• Lucky Loser

  

  Texas
  United States
  Member #86,149
  January 30, 2010
  1,891 Posts
  Offline
  Yesterday, 11:23 am

  ▲▼

  I haven't been here in quite a while, outside of checking results, so  I decided to see what the conversations looked like. This really piqued my attention because I initiated very similar discussions years ago regarding lottery draw processes. You eluded to the term 'random' and my personal position on what lotteries do, be it mechanical or RNG software, is the farthest thing from being random. I'm going to search for a link to my topic as a reference to my argument(s) which drew great input from members at that time. Some supported my theories while others felt that it made no difference on the official results. However, I put the work in and demonstrated in real time how pretests and post tests directly affect what one member said was the 'natural flow of numbers' which I agreed with her 100%. I like your topic and will be back to rehash my position(s).

  Small games, frequent wins, and regular payouts 'cause.....

  There are seven days in the week...'Someday' isn't one of them.

  #lotto-4-a-living
• hypersoniq

  

  Pennsylvania
  United States
  Member #1,340
  April 6, 2003
  2,942 Posts
  Offline
  Yesterday, 12:40 pm

  ▲▼

  Markov chains seemed like an interesting approach. I put this to followers as that seemed to be the most straight forward application... get a distribution of the numbers (all of them) that followed the last draw and present the distribution list, sorted by frequency descending... this is literally the layout of the transition probabilities for each state transition (draw). Even though, like the lottery, Markov Chains are "memoryless" and only depend on the last state, lottery data does not even do this.... having used chi square goodness of fit on samples ranging from 10 draws to entire histories, the scoring indicates that the groups are indeed representative of discrete uniform distributions. Bottom line, not enough winners predicted to make a system profitable.

  I also did that inter quartile range of Hot/Neutral/Cold... the problem was "when" was this hot? What sample size was the right one to use? To overcome stale hots... I added a count since each number was last drawn. The sample size chosen was the number of draws for a number to appear 15 times in the sample vs the expectancy... so pick 3, 150 draws could produce a frequency of 15, so how much above it or below it determined it's hot cold or neutral label.... then I refined it to include neutrals above, exactly at and below the median, giving Hot, Neutral Above median, neutral median, neutral Below median and Cold. None of it helps predict anything.

  I am taking a break to learn bioinformatics and biostatistics, but once I get comfortable with the concepts and techniques of Hidden Markov Models, I may revisit the analysis of lottery data through that lens... but right now, out of ideas.

  Have an EXCELlent day!
• NumberCruncher2

  

  us
  Member #261,885
  June 21, 2026
  16 Posts
  Online
  Yesterday, 2:46 pm

  ▲▼

  This is exactly the kind of rigor I was hoping someone here had actually done. The chi-square goodness-of-fit across sample sizes is the part most people skip- they eyeball a frequency chart and call something "hot" without ever testing it against a uniform null. The fact that it kept reading as discrete uniform from 10 draws up to full history is the whole answer, really.

  Your IQR refinement is clever - adding the "count since last drawn" to deal with stale hots is the kind of thing that sounds like it should rescue the signal, and I'm not surprised even that landed on "none of it helps predict." That's almost the most honest result you can get: you built the careful version and it still came up uniform.

  The Hidden Markov Models angle is interesting though - a latent-state model is at least a different question than the memoryless transition matrix, even if I'd bet the draws stay stubbornly independent. If you do revisit it through the bioinformatics lens, I'd genuinely want to hear what you find. Good luck with the biostats - that's a deep well.