Value Predictor

[stan]

[quote=Maryland]
After reading the papers on the value predictor it seems to say I should use less data rather than more for a good output. Does anyone think I'm interpeting this wrong?
[/quote]

the new algorithm selects only one historical pattern that matches best the latest pattern. while the previous algorithm (trend predictor plugin) used several matching historical patterns to make the prediction so more data was needed.

[stan]

[quote=stan:1154417345]
[quote=Maryland]
After reading the papers on the value predictor it seems to say I should use less data rather than more for a good output. Does anyone think I'm interpeting this wrong?
[/quote]

the new algorithm selects only one historical pattern that matches best the latest pattern. while the previous algorithm (trend predictor plugin) used several matching historical patterns to make the prediction so more data was needed.
[/quote]

btw, the algorithm searches for the best match from the latest data to the oldest (i.e. backwards) so even if you have a lot of data in the input and there's more than one historical pattern that matches the latest data, the latest matching pattern will always be used for the prediction

[Maryland]

Thanks Stan, and I know your "much" meant match

[stan]

[quote=Maryland]
Thanks Stan, and I know your "much" meant match
[/quote]

i fixed the typo, thanks

[Midas]

Stan,

For optimised performance of the pattern matching algorithm, would you include the following tuneable user parameters.

(1) Under user settings, provide an entry for "Min Pattern Size". This will allow user to search any range of pattern sizes(k). Default =1.

(2) Under user settings , provide an entry for optimise weights, call it "Sensitivity" range [0.1-1].This will allow the user to adjust the sensitivity of the historical pattern matching offset weight (wi). Default =1.

(3) To address some of Padawan's query on value prediction, here is a useful graphical trend aid.

(a) Provide a right hand side chart on the user setting page, showing the actual trend value as per summary chart.

(b) Overlay a long term moving average( 10 point say) with a short term moving average (3 point say) on the same graph.This technique for trend identification is based on the location of the shorter moving average relative to the longer moving average. If the shorter moving average is above the longer moving average, the trend is considered up. If the shorter moving average is below the longer moving average, the trend is considered down. If possible both moving averages could be adjustable, and the gap area difference for increasing/decreasing in different colours.

Cheerz
George

[stan]

[quote=George]
For optimised performance of the pattern matching algorithm, would you include the following tuneable user parameters.

(1) Under user settings, provide an entry for "Min Pattern Size". This will allow user to search any range of pattern sizes(k). Default =1.
[/quote]
will do

(2) Under user settings , provide an entry for optimise weights, call it "Sensitivity" range [0.1-1].This will allow the user to adjust the sensitivity of the historical pattern matching offset weight (wi). Default =1.

will do
btw, i was thinking that the weights could be applied in increasing values, i.e. it's more important that the pattern 'ends' match as closely as possible. so if the pattern had e.g. 10 elements than the weights applied could be 0.1, 0.2, 0.3, ... , 0.9, 1.0

(3) To address some of Padawan's query on value prediction, here is a useful graphical trend aid.

(a) Provide a right hand side chart on the user setting page, showing the actual trend value as per summary chart.

i'd rather integrate the predictor into winning numbers statistics and wn history tables/charts - imagine a new 'prediction' row added after the latest draw in stats tables showing the prediction for most tracked parameters, there could also be an extra 'prediction' point in the charts.
if you think the predictor results are good enough, i can start working on it

(b) Overlay a long term moving average( 10 point say) with a short term moving average (3 point say) on the same graph.This technique for trend identification is based on the location of the shorter moving average relative to the longer moving average. If the shorter moving average is above the longer moving average, the trend is considered up. If the shorter moving average is below the longer moving average, the trend is considered down. If possible both moving averages could be adjustable, and the gap area difference for increasing/decreasing in different colours.

same answer as above i suppose. but can you email me psl some example of expert lotto chart with the trend lines above? - just to make sure we're on the same wavelength...

[Midas]

Stan,

Thank you for the feedback!.

Item(1) OK
Item(2) This looks very good much like an exponential weighting scheme, where most recent pattern offsets are given greater significance than past historical patterns. (I also have an enhancement for you shortly)
Item (3a) Agreed. Can we hold embedding until we complete testing of the enhanced beta algorithms. Do we also want to add the fourier noise filtration/injection scheme now or hold until later stage?.
Item (3b) Agreed. Please find below an example of the graphical trend tool.

These charts graph a sample segment of WN History (Sum-4) for the UK National Lottery (6/49).

Chart 1 - Value Predictor
Series 1= Area Chart of Sum-4 Table.
Series 2= Short Moving Average(SMA 3_point)
Series 3= Long Moving Average(LMA 7_point)

Chart 2 - Moving Average Differentials
Series 1 = Delta (SMA-LMA), Notice the turning points transposed back onto original curve.

Some thoughts, regarding the final version of this tool. The end points can also be shown with error bar candles of standard deviation for prediction.

There's plenty to consider and test as we move forward towards the final version.

Cheerz
George


[/]

[stan]

[quote=George]
Item (3a) Agreed. Can we hold embedding until we complete testing of the enhanced beta algorithms. Do we also want to add the fourier noise filtration/injection scheme now or hold until later stage?.
[/quote]
it doesn't really matter as only the predictor core will change during fine-tuning while the integrated user interface should remain the same.
but for the moment it's easier to make changes in the plugin and upload new plugin versions only.

fourier noise would be a nice excercise but it will take me some time to implement it. it may be difficult to test it as well but i'll try

Some thoughts, regarding the final version of this tool. The end points can also be shown with error bar candles of standard deviation for prediction.

error bar candles ??

and thanks for the charts, i suppose they could be applied to other expert lotto charts as well

[stan]

now you can download value predictor version 1.0.1:
http://www.expertlotto.com/files/previe ... 0.1.plugin

it has new options in the settings tab:
- min pattern size - should be selfexplanatory
- weight & weight step - this is the coefficient applied when comparing the patterns (see the algorithm description). weight step specifies whether the weight values are decreasing or not. so if you set weight step to e.g. 0.1 then the coefficients will be e.g. offset = 0.7*x1 + 0.8*x2 + 0.9*x3 + 1.0*x4 when pattern size is 4

[stan]

yet another version of the value predictor 1.0.2:
http://www.expertlotto.com/files/previe ... 0.2.plugin

now it provides an alternate prediction where the best pattern size is chosen from all the predictions that have the most valid predictions for the past xx validation row. it seems to me that these results are slightly better, try setting the 'number of rows for validation' to 10 or 15 only and keep experimenting

[Alcymart]

Hello George, I have tremendous experience in experimenting how to make predictions with Recurrence analysis, however it's not part of my suggestion as we don't want to recode the whole plugin. Instead we should add on and improve it.

Every method of prediction is good whether it's based on the chaos theory, fuzzy point, neural networks, you name it.

What we should do is improve on what exists already. This is the best way to not get in the way of copyrights, patents and what not.

I have discovered a few years ago that to improve upon a method, the software or plugin in this case must be able to evaluate it's own predictions. The trend predictor seems to somewhat fit in this category however whether it's the trend predictor or Value predictor which I both tested, their success rate leaves to be desired compared with the success I acheived in my work, and this must be due to the fact that it relys on past draws and not True future draws simply.

Here is how the prediction value evaluator should be coded:

We would select the date from where the scan should start. Once the date is entered, the end date would be entered automatically as it would leave out the last 5, 10 or 20 draws. The evaluator then engages and Predicts the values based on the number of past draws like it is coded right now basically.

The predicted numbers are based only on the portion of database, beginning with Starting Date and ending with Ending Date. The results left out of the scan are TRUE Future draws in relation to predicted values. By comparing each subsequent future Draw against the Predicted values, we can evaluate true winning/losing performance of the prediction.

As each game type is different, in relation to its mode of operation, type of Barrel and the number if Barrels used (rotated), type and number of numbered ball sets used (rotated) and other mechanical and atmospheric conditions, it is impossible to lay down any specific rules on the length of the database to use (number of draws scanned) to predict and evaluate predictions. However, in my experience, a good suggestion would be to start at the very beginning (first draw in database) but also leave us the choice to scan from a specific date because If the evaluation is no good, we start again, but this time we would take off the last 10-20 draws (if database is large enough) or 5-10 draws if Database is smaller.

And all this scanning until the prediction has acheived the highest success rate. It's not the holy grail, but sure beats hands down what we have now. It must be able to evaluate it's own predictions and output the best values based on it's evaluations of True future draws! Not just output values based on all draws, but True future draws you see. It makes a whole world of difference.

I wouldn't know how to code this exactly as all I know is a little C++ and that's it, but I am sure you can figure this out quickly as you know what's in the code right now and how it operates.

Best Regards,

Bernard St-Pierre

[stan]

[quote=Alcymart]
Here is how the prediction value evaluator should be coded:

We would select the date from where the scan should start. Once the date is entered, the end date would be entered automatically as it would leave out the last 5, 10 or 20 draws. The evaluator then engages and Predicts the values based on the number of past draws like it is coded right now basically.

The predicted numbers are based only on the portion of database, beginning with Starting Date and ending with Ending Date. The results left out of the scan are TRUE Future draws in relation to predicted values. By comparing each subsequent future Draw against the Predicted values, we can evaluate true winning/losing performance of the prediction.
[/quote]

this is how the value predictor works already. you can enter 'starting row' and 'number of rows for validation' in the settigs & results page. as of version 1.0.2 the second prediction result printed is the one with highest success rate for the xx latest rows.

[Alcymart]

I wasn't aware that it was coded this way already. George is already doing a great job then! We are happy to have you guys not only hard at work but hard at thinking too!

As long as we all keep our neurons working, we will improve predictions all the time. I will give this some thought as to there is a limit to what we can acheive prediction wise. I know this by experience, trust me, and efforts should be put somewhere where improvement is not saturated... but I guess that's how you guys work and think every day already.

I have nothing more to say other than... Great Software!!!

[stan]

value predictor version 1.0.4 is now available for download:

www.expertlotto.com/files/preview/value ... 0.4.plugin


what's new:
- fixed mse & mape calculations (thaks george)
- the predictor will pick the best weight value for you (see the algorithm description). just enter min/max values and the step and the predictor will try all weight values in range min - max and choose the best one. 'decrease step' is the previous 'step' so if you enter non-zero value the weights applied on each pattern will decrease by this value
- you can get some more details on the prediction results (see the check box at the bottom)
- you can run the predictor in back-test mode (see the check box at the bottom). if the back test rows is set to e.g. 15 then the predictor will make 15 predictions and print its results with the actual values. so effectively the back-test mode is the same as running 15 predictions manually while changing last rows to ignore from 1 to 15 and comparing prediction with the actual values from the input page. (note: if last rows to ignore is set to e.g. 10 and back test rows is set to e.g. 20 then the predictions will be made for rows 10 to 30 counted from the bottom of the table)

more updates wil follow

note: i noticed that selecting a small number of rows for validation yields better results for me. try values in the range of 5 to 10

[Maryland]

Should your input be as the Trend Predictor with lastest at bottom for the Value Predictor?