I started mechanical investing back in ’98 with a couple of other engineers. Went through Value Line, Stock Investor Pro, and GTR1, and started using P123 back in 2005. P123 as I have mentioned before evolved just like the mechanical investing board. Bothe started off with inexperienced individuals finding some unique small set of factors picked a small number of stocks that backtested well but seldom worked out of sample. Over the last 20+ years P123 has built a much more powerful set of tools and matured their access to their historical data base by allowing you to create formulas that evaluate how a factor for a stock has held up over time. For example:
@EBITDA2EV:FRank("((EBITDA(0,QTR)*8) + IsNA((EBITDA(1,QTR)*4),0) + (EBITDA(2,QTR)*2) + IsNa((EBITDA(3,QTR)*1),0)) / EV", #All, #Desc) >= 50
And if you don’t understand what that formula does with that factor just ask their AI agent what it does.
Alphanaut
• @EBITDA2EV: assigns a name/label to this rule (useful for referencing or readability). It doesn’t change the calculation.
• EBITDA(n, QTR) = EBITDA for quarter offset n:
o 0 most recent quarter, 1 prior quarter, 2 two quarters ago, 3 three quarters ago.
• The numerator is a weighted sum of the last 4 quarters of EBITDA:
o 8*EBITDA(0,QTR) + 4*EBITDA(1,QTR) + 2*EBITDA(2,QTR) + 1*EBITDA(3,QTR)
o This heavily emphasizes the most recent quarter.
• IsNA(expr, 0) / IsNa(expr, 0) replaces NA with 0 (here applied to offsets 1 and 3), preventing the whole expression from becoming NA if those quarters are missing.
• / EV divides by Enterprise Value, producing an EBITDA-to-EV style yield (but using that custom weighted EBITDA instead of TTM).
• FRank("formula", #All, #Desc) computes the percentile rank (0–100) of that ratio across all stocks in the universe, with higher values ranked better (#Desc).
• >= 50 keeps stocks in the top half of the universe by this EBITDA/EV measure (percentile 50 or higher).
By the way that is from a screen that does very well.

My first full AI system has was built using only factors that worked before ~2004 and waking forward year by year tested with out of sample data for one year, retrained with that extra year of data till 2021. From 2021 the system held up and consistently performed out of sample. I hadn’t funded it to this point because I have a different system that works (but not as well) and as one of those older guys and I like a little more assurance than most.
An example of an AI model I am planning on funding soon:
General Info

Number of Positions 20
Last Trades (7) 09/28/26
Period 09/27/21 - 09/27/26
Sizing Method Static Weight
Last Rebalanced (Every Week) 09/28/26
PIT Method - Prelim Use
Benchmark S&P 1500 Composite (SPTM:USA)
Universe S&P1500 CompositeCap (IVV+IJH+IJR)
Ranking System SP1500EnsemV5
Quick Stats as of 9/27/2026
Total 5yr Return 352.62%
Benchmark Return 82.91%
Active Return 269.70%
Annualized Return 35.28%
Annual Turnover 331.51%
Max Drawdown -22.99%
Benchmark Max Drawdown -24.14%
Overall Winners (219/412) 53.00%
Sharpe Ratio 1.40
Correlation with S&P 1500 Composite (SPTM:USA) 0.79
Same Model with the largest 2200 MktCap US stocks does 5% better over last 10 yrs but with more volatility.

P123 takes a wile to become familiar with if I were just beginning subscribe to one of the less expensive plans and spend some time experimenting. The Ultimate Plan only makes sense for large Portfolios and serious investors.