# Moontower #184

## GPT Stuff

A college buddy texted me:

The link referenced is totally đź¤Ż. My friend just scratched the surface of whatâ€™s possible!

Enjoy:

• How to Use ChatGPT on Google Sheets With GPT for Sheets and Docs

What Can You Do With GPT-Powered Google Sheets?

Step-by-step walkthrus in the article on how to do the following GPT functions en masse in a spreadsheet:

1. Generate Text

2. Translate Text

3. Summarize text

=GPT_SUMMARIZE(C44) will summarize the content of cell C44 into the active cell.

4. Extract data

#### GPT-4

OpenAi released GPT-4 this week. Here are some buzzworthy examples:

This thread includes a similar example as well as prompts asking GPT-4 to create videogames, a link about Khan Academy building on the tech to create an assistant for teachers, and more.

Iâ€™m far too stupid to pontificate on what any of this means. Every day substacks, articles, papers, interviews, and videos comment on AI, alignment, the meaning of creativity, and the future of jobs from writing to coding. I just see a useful tool to use until the day Iâ€™m deleted from the simulation in favor of a paperclip.

## Money Angle

In the past few days, Iâ€™ve been getting around to the feedback and follow-ups from last weekâ€™s StockSlam sessions. Hereâ€™s a reaction and my response worth sharing widely.

Attendee:

Just wanted to shoot you a quick note – loved the game last week, thanks for putting it on!

I had a hard time playing the game because I didnâ€™t have intuition for the odds of the gameâ€¦ Iâ€™m way more of a Quant- the only thing I could think of was trying to execute the optimal strategy.

To figure out the optimal strategy Iâ€™d run a Monte Carlo simulation – play the game 100,000 times (programmatically using python or something) and see the distribution of outcomes as well as figure out some conditional probabilities (like what are the odds of last place winning given current relative location). Getting a sense of this would help price different bets – not a sure thing all the time, but better odds!

Generally, I ended up playing the game buying out-of-the-money â€śhorsesâ€ť (i.e. last place)â€¦ I figured with the mean reversion built into the game combined with behavioral biases to dump losers would be a winning strategyâ€¦ and I ended up with a positive PnL so maybe I was into something!

I donâ€™t know how you did that for a career for so longâ€¦ so stressful and I was wound up all night from it, hahaâ€¦

An anecdotal observation â€” I’ve noticed that quants and accountants actually get a bit paralyzed sometimes and it highlights the fact that crunching the numbers to perfection isn’t the core skill of trading.

It really is handicapping how wrong you could be and then acting with a margin of safety commensurate with the possible reward. Basically, if you wait to have the best info you’ll be too late. So the constraint is â€śhow do I act optimally subject to being fast?â€ť Everyone is in the same boat. Thatâ€™s a key point. The game would be different if everyone had infinite time to crunch the numbers. Trading is playing the game at hand â€” and that has a speed component. This is inescapable. Itâ€™s also true in reality even if the form varies. Buffet might wait for a fat pitch, but when it comes the bat speed still needs to be fast.

Whatever your game, you ultimately get a feel for it by being able to hold your attention on what matters and tuning out the rest. There’s some visualization…being ready to pounce on an incorrect market that you’ve been studying. In StockSlam, you really get a sense of what consensus is for a color in a certain relative position and then your antennae is up for aberrations. You are gathering and measuring data via listening and memory while in real-life the same functions are performed in code. But they are the same functions. And both are downstream from â€śwhat do I need to be paying attention to?â€ť That will vary by the time horizon of your strategy.

[The attendee also mentioned that the penalty for not executing the gameâ€™s â€śbroker cardsâ€ť was too low.

My response:

As far as the penalty we are actually thinking to ditch it anyway and use carrots for doing things on your card rather than punishments. But I hear you on the \$5 not mattering much but it remains a useful part of the game by letting us examine if players can find the least expensive way to execute the card. You are effectively benchmarking a trade not to “does this have edge” but “is this better than negative \$5”.

This is a critical concept in real life. Broadly, satisficing is often better than making perfect the enemy of the good. Also, there are some strategies that are not profitable if you have to cross a spread but are profitable if the benchmark is “it saved me from crossing a spread” (very relevant for an org that has to make many hedging trades per day). Academic papers are notorious for finding strategies that underappreciate indirect transaction costs. But you may be able to repurpose such strategies to warehouse risks instead of crossing bid-asks to shed them. Thatâ€™s a lower bar than a strategy that needs to cross a spread. In a world of rebate liquidity this is especially true. The cost/rebate structures for taking/ supplying liquidity is like a 4-point swing in a basketball game.]

Related reading (as an exercise you can think of why these posts are so related to what I described above):

• If You Make Money Every Day, You’re Not Maximizing (28 min read)

## Money Angle For Masochists

76ers GM Daryl Morey is one of the pioneers who brought Moneyball-type thinking to basketball during his tenure with the Rockets.

His interview with Patrick on Invest Like The Best is insightful and entertaining. I want to zoom in something Morey says:

You are weighing championship odds. And generally, we look over a three year time horizon with that. You could really pick any time horizon, but three years seems to work best with the data. And we basically do a sharp ratio like you would in investing, which is like here’s how championship odds increase, here’s the variance of that move.

Is it on the efficient frontier of return to risk basically and Shane [Battier], obviously, fit that for us.

None of our information is anywhere as good as the financial models. Actually, our underlying data is more predictive, quite a bit predictive. I talk to a lot of quants on Wall Street, and I tell them our signal to noise ratio using whatever measure you wantâ€¦.And they go like — yes, they go like, whoa, you guys are — that’s incredible. And I’m like, yes, but you remember, we have to be best of 30. You guys just have to beat the S&P by 2% and you are geniuses. So each industry has its own challenges.

We’re like a pure play. It’s the lifeblood of our business, whereas in other businesses, I’d say execution probably matters a lot more. In all aspects, including coaching, a well-executed, slightly suboptimal strategy generally will be the best strategy poorly executed. I mean you know that.

That’s generally true in basketball as well. But I would say in our realm of decision-making, it’s really almost a pure decision-making thing. This draft pick beats that draft pick. This free agent for \$5 million beats that free agent for \$5 million. It’s more of a pure play.

Sports is actually way simpler than most of the people you talk to, way simpler. Our sport, it changes, but not much. Our data is pretty good. Our competitors aren’t coming out with new products. Our competitive dynamics are known.

They’re hard, but they’re — no, we don’t have the Rumsfeld problem of unknown unknowns, like some start-up in stealth mode that might emerge, like, that’s why academics have done more and more papers about sports.

Because if you’re trying to isolate how to make good decisions, sports is really the right area to do that in

This is a great section because it highlights how different domains just have different size error bars. Sports signals are stronger than investment signals. The counterbalance to that fact is when Morey says:

I talk to a lot of quants on Wall Street, and I tell them our signal to noise ratio using whatever measure you wantâ€¦.And they go like — yes, they go like, whoa, you guys are — that’s incredible. And I’m like, yes, but you remember, we have to be best of 30. You guys just have to beat the S&P by 2% and you are geniuses. So each industry has its own challenges.

Umm, beating the SP500 by 2% consistently is rarified air even if that number sounds small. Morey admits that only a handful of teams have the requisite talent to even compete for the title. So your probability of winning the championship is either 0 or likely much better than a professional fund manager beating the SP500 by 2%.

Asset managers win by being good salespeople (a friend called this the Matt Levine model â€” being a good hedge fund is about gathering assets when you get hot and keeping them when you get cold. Itâ€™s a scheme for getting rich that has a lot less to do with returns than the industry will admit. Come to think of it, being a valuable sports franchise probably has more to do with the logo and stadium than actually winningâ€¦itâ€™s not that winning and returns donâ€™t matter, itâ€™s the gap between how much they matter and how much we think they matter).

Iâ€™m guessing Morey threw the 2% number out there without much thought. He was actually making a deep point that if an adversarial game is technically easier (say checkers vs chess) the competition enjoys the same low-difficulty advantage and you are in the same place of having a low chance of winning. But I was curiousâ€¦how hard is it to beat the SP500 by 2%?

Iâ€™ll admit a question like this is in my friend Nick Maggiulliâ€™s wheelhouse so when he reads this heâ€™ll almost certainly have a more complete answer. But I decided to take a quick stab at it.

I pulled up the portfoliovisualizer.com fund screener and filtered for US equity large-cap funds with at least a 5-year history benchmarked to the SP500 total return (this is an appropriate benchmark for a large-cap US equity fund.)

My criteria for beating the SP500 without getting lucky was the fund needed an information ratio (IR) of .50 or greater. An information ratio is outperformance normalized by tracking error. Tracking error is the standard deviation of the difference in returns between the fund and the SP500. If a fund outperforms by 2% per year but the tracking error is 10% (ie an IR = .2) that feels like noise vs a fund that outperforms by 2% with only 4% tracking error [I realize Iâ€™m using a simple, satisficey method for separating signal from noise, so if you are an allocator who just threw up in their mouth, brush your teeth then email me with an education so I can learn too!].

What did the screen turn up?

• 45 out of 677 funds had IR of .5 or greater (caveat: the IRs use a 3-year lookback)
• 8 funds out of 677 had at least .5 IR AND outperformed the SP500 total 3-year returns by 200 bps
• Only 3 funds outperformed by 200 bps for 5 years (the IR ratio is still a 3- year lookback)

Daryl your point is well-taken but beating the SP500 by 2% with skill is 90s Bulls-level for public fund managers.

## From My Actual Life

My music school does a class where you form a band for 5 weeks then perform. Itâ€™s a great way to accelerate learning. Iâ€™m on guitar duty for the show tonight.

Hereâ€™s the setlist:

1. Canâ€™t Let Go by Robert Plant and Alison Krauss
2. Far From Any Road by The Handsome Family