How strong is the arbitrage effect of P123 members?


TL;DR: I think downloads are key. We can share general methods (to trim or not to trim data on stock returns for example), keep a few secrets if people are concerned. And pursue your own ideas without having to convince everyone else of its worth or how high of a priority it should be (which could actually be a big secret that you do not want to convince everyone else to start doing anyway).

Lets imagine that Yuval, Duckruck, Walter, SteveA, Azouz, pvdb, Jonpaul, Marco, Dan, Judith and others all started emailing each other after Yuval comes up with a unique method that he wants to discuss and perfect. Maybe it is so revolutionary, unique and effective that they even sign an NDA. Working together they automate it. But alas, despite the NDA, it leaks out to the rest of the community.

With the code that is published in the forum, Python and ChatGPT to answer questions for code -illiterates (like me) a large part of the community starts using it.

So, first of all, the idea that even the majority of the community would all agree that a method is great is pretty ridiculous. Even if it is great, I mean. The community is contrarian and egos prevent them from liking an idea that they did not originate. Not that some natural skepticism is necessarily a bad thing.

But in any case, would the method still work or would the benefit be arbitraged away once it became public at P123 and widely used by the members?

Obviously I do not know the answer. It does make me wonder sometimes about how much of my method I should discuss in the forum.

For example, do I want everyone at P123 being aware of the Day-of-the-week-effect and opening ports on Friday? Maybe switching over to TWAP for VWAP? Even if the idea is wrong do I really want everyone else doing it? Convince them to start doing some of the things that seem to be working for me?

And it is not like Whycliffes (originator of that thread on opening vs closing transactions) is ever going to thank me. Marco did thank me for my machine learning ideas in the past and does take some feature ideas from the forum (including some of mine). Thank you @marco for that!!!

BTW, @yuvaltaylor listened to my idea and acknowledged my contribution on opening vs closing prices. In addition, he is not in charge of product development so ultimately we don’t have to reach a consensus–or for each of us to even fully understand an idea—in order to get a feature request considered. Anyway, thank you @yuvaltaylor.

Even without specifics of the method would we want every Kaggle member coming to P123 and using advance Deep Learning methods and automated reinforcement learning online-methods to asses every factor and their interactions on AWS servers almost real-time (which would not be unreasonable at today’ prices and the right investment size)?

My take: The P123 community is not that large compared to say BlackRock who also invests in the small-cap, micro-cap space it seems. I am not really sure what Medallion is doing but they might not want to miss every clearly opportunity in the small-cap space. Others, including, BlackRock, Numrai etc., are probably already using the “super-secret” method. Maybe it won a contest at Quantopian and the person paying for the contest-prize noticed at the time. It might not even be new method.

Maybe anyone who visits Yahoo could figure out that opening prices thing. My only point being much of what we do is duplicated elsewhere, albeit much of it by professional fund (and even some retail investors). Screeners, at least, are not rare (e.g., at Zacks with some cross-over on analysts’ data I would guess).

And actually, there could be a benefit. To some extent what we do works because others think the same way we do (including the P123 members and BlackRock). You may have to get in earlier, but if you do get in early, the buying that comes after you get in could be a good thing.



The arbitrage effect in the P123 community isn’t going to be very large in my opinion, especially not for methodology. Methods of testing don’t get arbitraged; stock-picking strategies might, but to a very limited degree. That’s just my two cents, not gospel.


This question of arbitrage effect was actually one of the first questions I had when I joined. I also have not reached an answer.

It seems the entire field of factor investing has had a big effect on factor returns as we see many factors loose effectiveness over time. But I assume that is from a much much larger group of investors using them instead of just P123. Or maybe it is from other market conditions…

I think that answering this would be very difficult as you would likely need to see live portfolio performance before and after the “secrets” were shared. I don’t think you can use something like the core combinations as I assume many if not all of those factors are used in academic studies and thus are widely known even outside of P123. Also I don’t think it has live performance before the ranking system was public.

Maybe you can look at alpha decay of popular factors to see how fast they rise and fall as an estimate???

1 Like

First, we are not arbitraging. As Asness said, an expected return is not an arbitrage as it sometimes loses.

Second, back in 2006, the folks at P123 believed that institutional machine-learning violence would conquer small caps and cause P123 to lose its relevance, which didn’t actually happen.

Finally, setting aside for a moment that Medallion or other institutions aren’t all that amazing, there isn’t actually any evidence that some factors fail primarily due to being discovered, published, and used, rather than in-sample/pre-publication overfitting.

1 Like

My actual applied RS is not only publicly accessible but comes mostly from the default core RS and reduces the useless parts. Some of the others are from “small and micro focus” RS and PEAD factors added by me. Also, it is mostly EW.

Also, this is an RS that was publicized in another thread. Even though it was designed for Europe, it is still performing well in the US microcap universe so far in May after it was made public

20 holding sim:

5 holding sim:

There’s a very good argument to be made that price-to-book stopped working at around the same time that everyone started using it as a proxy for value. Once dozens of ETFs started to be based precisely on this ratio, how could it possibly work well any more? This is an extreme case, of course, but arbitrage is also the reason that Benjamin Graham’s net-nets strategy also stopped working, and maybe it’s even to blame for the O/S failure of Greenblatt’s “magic formula.” There was no price-to-sales ratio prior to the 1980s (it was invented by Ken Fisher in 1983) and it worked really well for a while. Everyone was using it, and it was greatly popularized by not only Fisher but also Jim O’Shaughnessy. And then it more or less stopped working, except during certain periods. Was it arbitraged away or not? I don’t know.

Because there are enough ETFs tracking negative value (“growth”) factors.

In addition, the historical returns of the net-net strategy have been grossly overstated by the practically untradable prices, backfill bias, reporting bias and survival bias of a number of stocks with low liquidity and market capitalization.

Using the approach used for many of these studies, which is to consider all stocks listed primarily in the US and rebalance monthly without consideration of transaction fees, we can even see incredible annualized returns of up to 2,086.04%.



Even if we limit the minimum market capitalization, as some papers do, the simulation result still has (fake) alpha.

That is, the Net-net policy may not have stopped working; it just never worked.

TL;DR: Stocks move on new information. That movement has a time-course. What else do you want to know to answer this question? Isn’t that what we mean by arbitrage in this context?

I do think arbitrage has a kind of esoteric meaning and gets in the way of this discussion. We are really trying to answer what new information moves markets and how quickly that information moves the market here in 2023. is alternative data disseminated as quickly? Maybe people buy credit card data because it take a while for retrial investors to get that information.

Some at P123 place important on finding new factors that make sense to them but are not being used by others (the equivalent of Fisher’s price to sales). We can argue it that really worked for Fisher or not (or O’shaughnessey) but some believe in that.

Even, I who likes the method more than the factors understand you have to have good factors. And the stocks need some turnover—in part because the new information eventually gets priced-in.

We have a post showing that for some models some of that movement can occur in the first day of trading on a Monday. "Next open" always the best - #47 by yuvaltaylor

Which just suggests this arbitrage can occur relatively quickly at times.

With reduced movement by Friday for some ports (on the now 5-day old information). I.e., The new information can be priced-in by Friday.

But why do we ever even sell a position? Isn’t it because, at least to some extent, the new information in our ranking system evenytually gets arbitraged way for basically every port ever created at P123?

I am not sure what more you would need to show that there CAN be arbitrage.

There has to be some arbitrage somewhere, at some point for the market to even work, I think.


Arbitrage requires that the asset that is long and the asset that is short are close enough together. For example, long APE and then short AMC is arbitrage, but long value stocks and then short glamour stocks is not arbitrage.

Correct. So it is a matter of not using the same definitions in the discussion, I think.

Unless you think that your discussion of different factors above comes under the umbrella of “arbitrage.” In which case, some people are overthinking this.

New information eventually gets priced-in. Full stop. The only question worth considering is: How fast and which factors work best here in 2023.?

The issue isn’t pricing, it’s about return prediction. The “pricing” narrative presupposes that efficient market theory is primarily valid, even though this is incorrect. By the terms of this common fallacy in financial scholarship like “information”, “reaction”, “arbitrage” and so on, it seems that speculators are just bees circling around fair value and will eventually help companies return to that fair value. In reality, it is more predominantly about taking advantage of pump-and-dump schemes for profit, as appears in PFIE.

Pricing emphasizes chasing a static 0th-order moment (price), while return prediction emphasizes computing unbiased estimates of the 1st-order moment (return). In optimization, we also need to consider the 2nd-order moment (variance), the 3rd-order moment (skewness), and the 4th-order moment (kurtosis) to mitigate risk.

I’m not interested in a lot academic discussion involving definitions.

Personally, I find it useful to have recent information. The more recent the better. For my present models and for my machine learning models that I rebalance on Fridays, I like to have the data from the overnight update. I find it preferable to the data from the previous weekend.

That has been my major thrust in the forum for a while.

Simple enough for me.

1 Like

I tend to rebalance daily, have you tested if this works for you?


Great question!!! My simple answer is not enough to reach a good conclusion that would be helpful to you.

I paper traded a port for a time. Seemed like forever at the time but not very long really. And I think I was not using good factors then either. So an n of one and a pretty lame n at that. But I was unimpressed with the results if that helps at all.


In my simulation, daily rebalancing increases annualized returns, have you tested the difference?

I think the only way that I have tested the difference is with the paper-traded port I mentioned above.

So I you have been able to backtest daily rebalance you are using largely technical data that is updated daily for daily rebalance at P123?

Interesting. Than you for sharing.



The value-based factors are also subject to change due to price data.

I’ve always assumed that the fundamental data is only updated on weekends in the simulation as well, isn’t that the case?


Correct on both counts I think. Fundamentals are updated weekly, I believe.

But you write formulas that update value-based factors daily based on price (or other technical data)?

So just trying to understand. If I use a FCF/Price as a pre-built P123 factor in a ranking system it might be updated weekly in a sim ranking system.

But you thought of putting FCF/close(0) in a by rule?

Not sure, but very smart if that is what you are doing to get around that.

Edit: And one could kind of simulate the ranking system in a buy rule like: 1 * FCF/close + 3 EBITDA/close(0)……Using the normalized weights of value factors that include price in your ranking system? You might even be able to get a factor that is sort of like enterprise value using close(0). I am sure Yuval would know how to do that.

Second edit. You could even add weekday != 2 to your buy or sell rule so that your port uses the ranking system on Monday. You would have to include something so that you did not get wip-sawed selling on a Monday and buying on Tuesday with a Buy rule that was not exactly like the ranking system.

Maybe include a minimum price-change rule to make sure that a price-change is triggering the buy or sell along with a minimum holding period after being bought perhaps.

So like sell rule (close(0) - close(3))/FCF > 0.2; if you are pretty sure that would kick it out of your ranking system that included FCF/P. So it would not be bought again on a Monday unless the price had changed again (or a new FCF with an earnings report).


Just my .02 as someone who self admittedly knows nothing…

If you look as to why someone “deserves” alpha, it would be an informational advantage, a systemic advantage or a behavioral advantage. I think it makes sense that net-nets and Price to Book should be arbitraged away. First of all, these are simple single factor or formula arbitrages. To discover a net-net in the days of Ben Graham, it took a tremendous amount of work. You had to spend hours and hours manually piling through 13Fs looking for them. An investor should have been rewarded for that kind of informational advantage. It makes sense that that advantage should disappear when anyone can run a simple screen looking for net-nets and get a list of them in seconds on any number of free screeners provided by any large brokerage Likewise, I think it make sense that Price to Book should largely lose efficacy when anyone can pull up Vanguard and buy an Value ETF of the cheapest price to book with billions and billions in AUM in a matter of seconds. What advantage have they earned to deserve alpha?

I’m personally skeptical that P123 community itself could arbitrage away alpha like that using a multi-factor , even on small and microcaps. There’s just too many frictions involved, even if you were to just blindly follow the P123 Core system. It’s just too hard to implement on an institutional scale, and on the retail scale there’s a pretty sizable learning curve involved just to blindly copy an existing system, much less having the behavioral discipline to stick with it during underperformance. P123 users are still rewarded for informational, systemic and behavior advantages that come with being the type of person who becomes an P123 user. Now, if someone finds a way to create an ETF with a P123 microcap multi-factor system where anyone could emulate with a few clicks on Schwab or Fidelity, that’s another matter. But I personally think that would be difficult to pull off.

Without the help of long history, and before the current drawdown, investors might have mistaken value investing as safe.

  • Back in 2006, which is when systematic value investing started to go mainstream, investors could see just one value crash of -54% back in 1932.
  • It was easy to dismiss it as part of a ‘very different history’ that was no longer relevant. Investors made a similar incorrect assumption about Price Momentum before it crashed in 2009.
  • With the help of extended history, the hypothetical 2006 value investor, warned about the periodic crashes, would estimate value’s risk differently, and as a result, would be less hurt during the current drawdown (I think it’s time for Excel to recognize dates before 1900
    1940 to 2006 was an exceptionally safe period for value investing compared to its full history. And this tailwind helped many great value investors, creating unrealistically high expectations for value investing.
  • For example, the Graham-Newman partnership years only saw the lowest value drawdown of -40% in 1939 - just three years after their partnership started. From then on, value rallied with minimal drawdowns during the twenty years that they operated with astonishing 20% per year returns (13% of which could be attributed to the top decile of book to value Fama-French portfolio).
    • Warren Buffett was also a benefactor of the 50-year low-crash-risk period that ended around 2006. He is now navigating the worst value drawdown in his very long investment career.