The recommendation was simple: start with AI Factor right away

[ustonapc]

Judge and VS,

I saw these posts on X.

It seems both of you are doing better than the quant hedge funds YTD 2026 mentioned in the latest Bloomberg article below.

Can you share some more info about the P123 AI factor strategies and whether it is possible to subscribe to these? Based on the screenshots, it seems possible for P123 AIfactor to beat the best quant hedge funds (at least in the past 2 weeks.)

Thank you for sharing in advance.

Regards

James

Judge

My Long/Short (Substack: Systematic AI Investing Portfolios, see my profile) book is holding up well: Long AI Factor S&P 500 models, hedged with a hefty 25% short on the Russell 2000 (TWM). This is working even as IWM shows strength right now.

VS

One exciting part of strategy development is knowing when to push the pedal and then seeing the pedal push paid off. On new years I deployed a set of changes to my @P123Finance to better align with my utility function. The resulting changes = 50% > YTD 2026 returns.

Quants in Worst Losses Since October as Crowded Bets Buckle

By Justina Lee

January 21, 2026 at 9:11 PM GMT+8

Updated on

January 22, 2026 at 12:19 AM GMT+8

Takeaways by Bloomberg AI

• Quant hedge funds are off to a poor start in 2026 due to setbacks in US stocks, with losses reaching 1% in the first 10 days of January.

• The losses were concentrated in US equities, with US quants estimated to have lost 2.8% over the first two weeks of 2026, according to UBS Group AG.

• The recent selloff stemmed from factors including losses in crowded positions, short positions on high-beta names, and adverse idiosyncratic moves, according to Goldman Sachs.

Quant hedge funds are kicking off the year in the red, as setbacks in crowded US stocks clobbered the strategies, reviving concerns about volatile returns in the sector.

Early January marked the worst 10-day period for systematic long-short equity managers since October, with losses reaching 1%, according to prime brokerage data from Goldman Sachs Group Inc. Most of the pain was concentrated in US equities, Goldman’s Kartik Singhal and Marco Laicini wrote in a note, drawing parallels with the sharp drops that hurt quant portfolios in June and July last year.

Losses at US quants amounted to 2.8% over the first two weeks of 2026, UBS Group AG estimated, based on its prime book. It noted that Friday had seen the sharpest one-day deleveraging since Dec. 22.

Market moves spurred by US policies this week may have taken some pressure off, but the damage has already been done. It remains to be seen whether systematic hedge funds will be able to claw back some gains.

Many quant funds ended 2025 in the black. However, they endured two violent loss-making periods, the first in early summer and then in October, when they were tripped up by reversals in momentum, coupled with a junk rally. Similarly, the most recent selloff stemmed from three main factors, according to Goldman: losses in crowded positions, short positions on high-beta names, and adverse idiosyncratic moves.

“The idiosyncratic drag has been driven predominantly by the short book, similar to the June-July drawdown,” Singhal and Laicini told clients, adding that this time momentum strategies helped cushion the blow.

Quants’ poor start to 2026 coincides with volatility on world markets. While AI and economic confidence fueled a rotation into riskier shares and small-caps when the year began, risk sentiment soured early this week amid President Donald Trump’s insistence on taking control of Greenland and renewed trade-war fears.

Since computer-based hedge funds typically rely on a variety of proprietary trading signals, it’s hard to pinpoint the exact source of the losses. Analysts taking a bird’s eye view often dissect performance based on so-called factors, effectively, stock characteristics documented by academia to drive returns.

The three episodes of pain, including the one this month, have one thing in common: a rally in so-called junk assets that saw riskier, more volatile shares surge. Such episodes tend to hit quants that typically short such low-quality companies.

“Although many quants try to be ‘factor neutral,’ it’s not easy to be completely factor neutral, especially when factors have significant moves,” Yin Luo, a quant analyst at Wolfe Research LLC, said in emailed comments. “Year-to-date, most factors are in the ‘wrong’ direction.”

Tuesday’s drastic risk-off shift might have helped ease the pain, added Luo, who noted that long bets popular among hedge funds rose, while heavily shorted names fell.

But even after taking into account those gains Tuesday, a strategy that goes long steadier shares and short the opposite has lost about 4.3% so far this year, while the ‘quality’ trade that favors more profitable, less leveraged stocks has dropped around 0.8%, S&P Dow Jones indexes based on US stocks show.

Global stock markets recovered Wednesday after Trump said the US doesn’t want to use excessive force to get Greenland.

“When you go from 0 to down 5 in the first couple weeks of the year, you’re getting a call from the risk manager, and it’s not to wish you a Happy New Year,” said Blago Baychev, co-founder of PharVision Advisers, a systematic equity hedge fund. “Crowding, both on the long and the short side, is the ever-elusive risk factor that everyone is trying to either hedge or dodge, but it is getting increasingly difficult.”

(Updates market moves throughout, adds Wednesday market context in penultimate paragraph and manager comment in last paragraph)

[judgetrade]

Dear James, I send you a DM.
Thank you and
best Regards
Andreas

[VSai23]

Hey James. Sent you a message as well.

[ustonapc]

Thanks, I have replied the message.

[pitmaster]

I see no clear evidence that ML-based Designer Models have outperformed ranking systems (RS).

My view is that traditional RS excel during small-cap rallies (like the current one) because factor signals become strongly monotonic —higher ranks systematically yield higher returns. In these linear regimes, simple rankings capture alpha efficiently. Conversely, ML likely holds the advantage in sideways, hype-driven, markets by exploiting the non-linear, conditional relationships that linear models miss.For context, my live us micro-cap 20stcks RS strategy is up +23.58% YTD.

[Jrinne]

@pitmaster

I see no clear evidence that ML-based Designer Models have outperformed ranking systems (RS)

Some are using ML methods to optimize P123 classic ranking systems:

Genetic algorithm to replace manual optimizaton in P123 classic

These tools seem too complex to me and may increase the risk of overfitting to the data. If you prefer to focus on factor selection and economic intuition rather than factor weighting, you might consider a custom discrete optimizer. Such an optimizer can: Set factor weights on a fixed grid (e.g., min = 0.025, max = 0.10, step = 0.025). Accept a dictionary of fixed weights for core factors you always want in the model. The advantage is avoiding wasted computation on meaningless precision — …

ML Script Performance vs. Genetic Algorithms & CMA-ES**

I've been experimenting extensively with various approaches to build a machine learning model using offline data from the "factor download" feature, but for some reason, I'm struggling to achieve strong results. Below is my latest script, and I'm not quite sure what I'm doing that's preventing it from working well. So far, I'm getting decidedly better results—both in out-of-sample and sub-universe tests—by using a hybrid approach: Genetic Algorithms for factor selection and CMA-ES algorithms fo…

And most recently:

PCA Formula Exporter: Automatically turn Principal Components into Ranking Systems

@Scifospace came with a brilliant idea to break down the result from a Principal Component Analysis (PCA) Components to new factors for P123. While PCA is excellent for uncovering the "true" underlying drivers of stock returns by reducing multicollinearity and distilling correlated factors into independent components, translating the math into Portfolio123 formulas is incredibly tedious. To make this effortless, I built a small desktop application: PCA Formula Exporter. What it does: It takes…

P123 Classic has always been a linear GAM (Generalized Additive Model) by definition. We have moved from mostly manual optimization to what is clearly a gray area for everyone—including those most against machine learning.

Going forward, I think some non-linear GAMs hold promise, not just for their transparency, but for their ability to handle non-linear monotonic relationships. This would essentially be a 'non-linear P123 Classic.'

Glass Box or Black Box: Explainable boosting machines let you choose

Prefer a glass box over a black box? Microsoft has been working on that. Here is a video: The Science Behind InterpretML: Explainable Boosting Machine - Microsoft Research "InterpretML is essentially a high-performance Generalized Additive Model (GAM) that uses Gradient Boosting to learn the non-linear functions for each feature. For P123 each feature could be part of a GAM with how each feature contributes fully visualized. It is an open source Python Program.

More and more, the question shouldn't be 'P123 vs. Machine Learning.' Rather, the question is: 'How can I use machine learning to optimize P123 Classic?' Even the use of Mod() (or odd/even universes) can be seen as a form of cross-validation.

From a marketing perspective, P123 could do more to frame this less as a debate between opposing sides. Maybe P123’s AI 2.0 will incorporate apps from @AlgoMan and others to help bridge that gap.

As it is, whenever someone likes P123 Classic, there is often an implied dislike for P123’s AI. This implies a divide that isn't good for marketing or the community.

P123 would do well—marketing-wise—to facilitate more ML options specifically for optimizing P123 Classic. This would soften the debate that often ends in implicit criticism of the AI tools, while providing tools for everyone. As I understand it, ML has helped the bottom line despite the negative sentiment in the forum.

[ars7777]

I agree there’s a duality in many of these topics — and maybe there doesn’t have to be.
Things that appear opposite are often connected.

As a bit of an outsider to this world, I’ve found it interesting to see the range of perspectives. I felt something similar in the market-timing discussions. It surprised me to see people who strongly believe in their ability to time which stocks to own reject the idea of timing when to be in the market at all.

I feel the same way about stop losses. In a 20–30 stock microcap system, letting rankings dictate decisions may well be the wiser path, and stops may be impractical in thin names. But more broadly, the statistical logic of embracing small losses makes intuitive and financial sense, and I assume Falnu’s model found a way to incorporate that.

For many of these debates — traditional vs. AI included — it seems less like “either/or” and more like what Jim Collins called the “genius of the and” being superior to the “duality of the or.”

[judgetrade]

Agree 100%.

The most important rule is to use what works for you. If your traditional ranking systems are performing superbly and AI Factor models aren't beating them, there's absolutely no reason to switch.

In my case, I was able to create superior models using AI Factor. That's why I use them, and the process has also given me a much deeper understanding of how to build robust traditional ranking systems.

[yuvaltaylor]

@Jrinne

More and more, the question shouldn't be 'P123 vs. Machine Learning.' Rather, the question is: 'How can I use machine learning to optimize P123 Classic?'

I agree. I have no idea how to answer that question, though. Everything in ML looks so difficult to me, and there doesn't seem to be much in common between ML validation and ranking system validation, at least the way I practice it. ML validation uses the entire spectrum of returns while ranking system validation looks at just the top. ML validation uses fixed holding periods and no slippage while simulations (the best way to validate ranking systems) take slippage into account and use flexible holding periods and complex buy and sell rules. ML has no way of dealing with the effects of position sizing by rank, while simulations deal with it pretty well. Maybe I'm just too stuck in my ways, but I really AM open to the possibility of using ML to optimize ranking systems. If it presented something close to what I want I'd jump at the chance to use it.

@ars7777

I felt something similar in the market-timing discussions. It surprised me to see people who strongly believe in their ability to time which stocks to own reject the idea of timing when to be in the market at all.

I wrote an article about this very conundrum, which you might find entertaining: Market Timing, Tactical Asset Allocation, and Trading: A Dialogue – Fieldsong Investments

@ars7777

he statistical logic of embracing small losses makes intuitive and financial sense

I'd like to hear more about this, please.

[Jrinne]

@yuvaltaylor

here doesn't seem to be much in common between ML validation and ranking system validation,

For example, the use of Mod() is exactly the same as Sklearn’s cross-validation with shuffle = True, but Mod() is limited. Providing additional cross-validation options within P123 Classic would be one way to start bridging the apparent gap between machine learning and P123 Classic.

Personally, I use P123 Classic for rebalancing. I also use a variety of Python tools as well as some basic math principles to optimize. I see value in a wide range of methods and am not sure which camp someone would want to put me in (ML, math geek, or user of P123 Classic’s ranking system).

I feel free to use whatever works. I agree that P123 Classic can be a very useful tool but I do not limit myself to using it alone.

[ars7777]

Your article is thoughtful and I agree with most of it, especially the idea that real alpha comes from exploiting inefficiency (like microcaps), not from trying to forecast broad markets.

Where I think the argument stops a bit short is in equating timing with prediction. Many modern “timing” systems aren’t attempts to forecast what markets will do next, but to detect the current regime. Trend following, dual momentum, vol targeting, and cross-asset confirmation are less about predicting direction and more about identifying when conditions historically associated with large losses are already in place.

The goal also isn’t to beat buy-and-hold on CAGR, nor to be right all the time, but to improve the path — reducing drawdowns and smoothing the ride without wrecking the return. For someone living off a portfolio, a system that comes close on return while cutting drawdowns meaningfully is very valuable.

That ties into what I meant by the “statistical logic of embracing small losses.”

In trading, an occasional large winner can pay for many losers if the losers are consistently small and controlled. For example, in my previous life, I would have recently taken SLV around $28 with a stop near $24 — roughly 15% stop. In say a $1M portfolio, risking 0.5% ($5,000) would mean buying only about $35,000 worth. Many times that trade would stop out and lose $5,000. That’s the small, acceptable loss. But occasionally it doesn’t stop out and instead runs far enough that it returns many multiples of the risk (as it would have in this case). The only way to ever collect those outsized moves is to be completely comfortable taking the repeated small hits that often precede them.

The same logic applies to the SPX put hedge. It consistently loses money, by design. But by accepting those small, ongoing losses, I can maintain much higher offensive exposure than I otherwise could, while knowing that in a real stress event I’ll have protection and deployable cash exactly when it’s most valuable.

So I agree that using timing to create alpha in efficient markets is likely a dead end. But using regime detection and risk overlays to protect alpha generated elsewhere feels very similar to the logic of stops and hedges: embrace small losses so the big gains can matter.

Edit: My main point is that if Falnu found a way to improve his system by intelligently truncating losses, that wouldn’t surprise me at all. One of the real strengths of a group as talented and well-rounded as this one is the opportunity to learn from each other’s approaches and see where we might refine our own.

[SZ]

I am always trying to learn new and better things. This year I have been learning about many “investors” in the US Stock Investing Championships and their methods and some AI ML items. Very different from what I do. It does not mean I will adopt all approaches, but if they work they might help you learn about the market as a whole or about a particular type of stock or situation. The way I see it you can always do both an AI model and a traditional model and take ideas from each. I have not had time to actually create an AI factor system though as emerging market investing has been quite active over the last 12 months for me. I could have made time but decided to improve my traditional systems while the pan is hot

EM data might be coming at a good time if this continues. Very interested in the upcoming addition

[judgetrade]

As a possible first step, if you use a trad. multi factor ranking system with a lot of factors –> simply use the factors as features (put all to “Skip and Date”).

Non AI Ranking System (OOS Live since the beginning of 2025), so doing well!

Copied 1:1 all 199 factors of the trad. ranking system to AI Factor Model, Predictor trained to 2020.06

All other settings are here:

HPs of the Predictor:

{"n_estimators": 500, "max_depth": 6, "learning_rate": 0.01, "gamma": 0.2, "booster": "gbtree"}

[SZ]

Thanks for sharing. Have been thinking about this. Good to see it working well! Looking forward to trying this out. Have you found more success with lightgbm or xgb?

[judgetrade]

XGB very new to me, usually I use ExtraTrees and LightGBM, but so far o.k., XGB does well with interaction features like

HighPct(63) * TRSD20D
ChaikinMFP(21,120,0) * (Close(0)/SMA(50) - 1)
(Close(0)/ChaikinTrend(45,0,5)) * (AvgVol(10)/AvgVol(125) - 1)
Rel%Chg_D(21) * Beta1Y

I mix them, so I would have one AI Factor Model –> produce one predictor with LightGBM one with ExtraTrees and have a look at the stocks it picks, usually it picks different stocks, so combining both portfolio strategies in one book gives good results. So XGB basically as well –> find more capacity.

Also tested for small cap models –> use lightgbm in the ranking system of the portfolio strategy.
Build the second ranking system with Extratrees and then add a buy rule to the portfolio strategy that uses the lightgbm ranking system.

Buy rule –> rating("3MRELReturn 30 Mil NU ExtraTOF") > 80

[SZ]

I see- interesting. Will definitely be trying a few combinations and ensembles. Thanks for the guidance

[judgetrade]

I made a mistake –> sorry!
the model I used for the predictor is a Zscore Date Model

[judgetrade]

and here is the universe both strategies use (the one with the trad. ranking and the AI Factor Model)

I go up to 30 Mil (e.g. below 30 Mil) + on some models I exclude Financials either on the universe or on the portfolio strategy.

[SZ]

Oh wow quite small! Good to know on the settings😬

[yuvaltaylor]

@ars7777

Many modern “timing” systems aren’t attempts to forecast what markets will do next, but to detect the current regime.

This is an altogether laudable goal, and very useful too--if they work. The questions then become:
a) How much of a lookback period do you need to "detect the current regime"?
b) Does that lookback period give you sufficient time to act before the "current regime" is over?
c) If you don't react fast enough to a change in regime, will that hurt your returns more than just applying an agnostic method that's non-reactive to regime changes?

A lot depends on how you define regimes. I'm hoping to investigate "factor reversal" regimes to find out a bit how this works in practice. Does anyone have any answers to the above three questions?

Also see: Change Partners: Some Thoughts on Market Regimes - Portfolio123 Blog