←
Richard Pham286 posts
Richard Pham avatar

Richard Pham

@phamrich_

cpo @fundamental

◷ Joined November 2022▧ 17 Photos and videos
Jeffrey Liu avatar
⟳ Richard Pham retweeted
Jeffrey Liu@jeffreyliu_05·Nov 17

We're excited to announce Catalyst Bench, the first evaluation to measure whether AI agents can generate excess returns with fundamental investing, and find that Sonnet 5.5 agents can achieve fundamental alpha by researching and modeling catalysts

Can AI agents generate fundamental alpha? Agents are now used in almost every part of fundamental investing, but it's not clear if their work can actually generate alpha In this new blog post, we asked agents to trade 320 stocks weekly over 13 weeks post their knowledge cutoff in Catalyst Bench v0.5 Agents researched and modeled business fundamentals using point-in-time data We show that Claude Sonnet 5.5 agents (across different harnesses) can generate alpha Sonnet agents beat 99% of random portfolios and nearly all technical strategies, and generate excess returns after correcting for simple factors In comparison, the weaker GPT-6 Luna agents aren't able to generate alpha There have been many previous attempts at measuring whether AI can make money in the stock market, most suffer from small sample sizes, so no real conclusions could be drawn This work shows early evidence that the latest frontier LLMs are powerful enough to be alpha-generative fundamentalresearchlabs.com/…

Richard Pham avatar
Richard Pham@phamrich_·Nov 17

getting more accurate estimates is one thing this is how we had agents generate Alpha

Can AI agents generate fundamental alpha? Agents are now used in almost every part of fundamental investing, but it's not clear if their work can actually generate alpha In this new blog post, we asked agents to trade 320 stocks weekly over 13 weeks post their knowledge cutoff in Catalyst Bench v0.5 Agents researched and modeled business fundamentals using point-in-time data We show that Claude Sonnet 5.5 agents (across different harnesses) can generate alpha Sonnet agents beat 99% of random portfolios and nearly all technical strategies, and generate excess returns after correcting for simple factors In comparison, the weaker GPT-6 Luna agents aren't able to generate alpha There have been many previous attempts at measuring whether AI can make money in the stock market, most suffer from small sample sizes, so no real conclusions could be drawn This work shows early evidence that the latest frontier LLMs are powerful enough to be alpha-generative fundamentalresearchlabs.com/…

Richard Pham avatar
Richard Pham@phamrich_·Nov 12

another look at our man vs machine: earnings predictions this is not luck or simple heuristics. these agents are reasoning about the fundamentals of the business, including how regional demand is impacted by hurricanes (1:36) now imagine 200 agents doing this for every line item in your financial model

here's how AI agents forecast earnings better than Street consensus by reasoning like a fundamental investor

2:41
Richard Pham avatar
Richard Pham@phamrich_·Nov 10

Our very own @jeffreyliu_05 did not receive formal finance training (In fact he did not complete college at all), but demonstrates remarkably high taste here in defining 'judgement calls'. Combine this with the true ML research expertise of him and @TianhangZhuzth and you can really start benchmarking your agent to the degree sophisticated investors need

Post attachment
Richard Pham avatar
Richard Pham@phamrich_·Nov 9

just, wow. Fable 5.2 saturating our benchmarks that include full financial model builds that takes humans >1 week gotta get back in the lab with our research team

Fable 5.2 is now available in Shortcut! Fable 5.2 sets a new accuracy record on our financial modeling benchmarks, scoring 91% at high effort versus 86% for Fable 5.1, and its low effort beats every other model's best score. Fable 5.2 has better judgement improvements in ambiguous situations and is better at inferring the user's intent from context. Fable 5.2 will be replacing Fable 5.1 in the model picker. Your organization's security protocols for Fable 5.1 carry over to Fable 5.2.

Post attachment
Richard Pham avatar
Richard Pham@phamrich_·Nov 9

are agents capable of fundamental company analysis that is more accurate than humans with decades of experience? we went deep on this TLDR; yes

Can AI forecast earnings better than Wall Street pros? In this new blog post, we show that AI agents can forecast earnings per share (EPS) better than the Wall Street consensus by reasoning about business fundamentals EPS is one of the most important metrics about public companies, and the Wall Street consensus comes from expert sell-side analysts But: the Street consensus systematically underestimates real EPS, so a simple heuristic like increasing the Street number by 5% would do well too Our key finding is that AI agents outperform the Street consensus by *understanding* the business fundamentals, and making adjustments to the financial models that are company-specific So not just a cheap heuristic I think this is some of the first serious evidence that AI can really understand company finance well enough to make decent forecasts Most of the work done by the fantastic @mishhman fundamentalresearchlabs.com/…

Robert Yang avatar
⟳ Richard Pham retweeted
Robert Yang@GuangyuRobert·Nov 4

went pretty deep with @FundamentEdge on how we went all the way from minecraft to investment agents our company pivoted big time from consumer entertainment to enterprise productivity, but really it's the same quest to understand the essence of a human and how to build it

Very fun to have @GuangyuRobert on the Invest With AI podcast to discuss the @Fundamental story and what we plan to build with The Digital Investor stack at Fundamental Research Labs. Inside the episode: - 1,000 AI agents in Minecraft that organized a search party on their own - Why going viral wasn’t enough to make Shortcut work - Why hedge fund analysts, not everyday Excel users, became the core customer - The case for one agent that manages your other 50 - Why today’s models fold the moment you push back - What it takes to build an agent that holds a real investment view TIMESTAMPS [00:00] Intro [00:43] Why Robert Wanted to Build AI Before It Was Popular [08:40] Running an MIT Lab and Questioning the Work [14:05] The Agent Research That Pushed Him to Leave [19:20] Project Sid: Building a Society of 1,000 Agents [22:18] The Search Party Nobody Programmed [26:58] From Minecraft to Excel Agents for Wall Street [33:05] Going Viral Wasn’t Enough: Making Shortcut Work [37:28] Why Hedge Fund Analysts Became the Customer [41:51] From an Excel Agent to an Investment Team [45:51] Why Nobody Wants to Manage 50 Agent Threads [50:12] How Close Are We to a Digital Hedge Fund Analyst? [53:52] The Problem with AI That Always Agrees [56:50] Teaching an Agent to Hold an Investment View [1:04:29] When Analysts Start Working More Like PMs [1:06:29] Brett’s Role: Training the Digital Analyst [1:09:07] If AI Frees Up Your Time, Where Does It Go? [1:11:34] What This Means for FE and Invest with AI youtube.com/watch?v=qVPJv5ga…

Richard Pham avatar
Richard Pham@phamrich_·Nov 2

if you're looking for a job and you care about having individual impact, revenue/person is a pretty strong signal quite remarkable how companies like Goldman are able to maintain this metric well into 7 figures, especially considering 3/4 of their staff aren't even in front-office revenue producing seats

revenue per employee tells you a lot about company culture & talent density jane street's revenue/employee only rivaled by anthropic ramp is highest among fintech startups, but still lower than goldman

Post attachment
Richard Pham avatar
Richard Pham@phamrich_·Oct 30

forecasting earnings better than humans DONE bringing alpha-generating insights to investors NEXT shortcut agents making fully autonomous investment decisions that beat the S&P returns...loading...

@tryshortcutai's agent has been on an amazing run over the last 22 earnings reports. Their earnings forecasts were closer to reported numbers than consensus 77% of the time. Really shows the power of a combination of domain expertise and deep AI agent engineering capability.

Post attachment
Richard Pham avatar
Richard Pham@phamrich_·Oct 23

accidentally said “improve” instead of “hillclimb” today and got kicked out of SF

Richard Pham avatar
Richard Pham@phamrich_·Oct 23

this just oozes rigor and craft. not even a fully baked product but it's everything I would want in an AI coach also @BrainsAndTennis post the full video of me cooking you

I grew up playing tennis competitively, and two years ago, before two kids, I was still pretty good (9 UTR). I’ve been getting back into it, but I wish I had a world-class coach telling me what to work on. But good coaches are expensive, selective, and hard to find, whereas models are cheap, available, and should be already good enough. So clearly, @nicochristie and I needed to explore what's possible. We chained open-source models post-trained on tennis. This cut each broadcast down to live play, found and tracked the players, ball, court, and labeled every stroke with its type and a 65-joint skeleton. We then post-processed this data into 30+ biomechanical traits across serves, forehands and backhands, normalized for each player’s build and the camera. This included the goat @DjokerNole, @carlosalcaraz, @DaniilMedwed, @FTiafoe, @AlexZverev, @BenShelton, @Taylor_Fritz97, @NickKyrgios, @AndreyRublev97, @janniksin, and @TommyPaul1. Using the same pipeline, we then processed phone footage of myself and @phamrich_ (ex-D1, used to be >13 UTR) playing. Then we gave Claude all the measurements with a blank context. What impressed me: - Without being told, it found my weaker side (forehand) and Richard’s (backhand) - It joined differences that looked separate into one root cause per stroke - My own diagnosis of my forehand was wrong. Claude’s makes a lot more sense What do you think? @bgtennisnation @paul_annacone @darren_cahill Drop youtube links here if you want Coach Claude to give you tips

0:16
Misha Boyko avatar
⟳ Richard Pham retweeted
Misha Boyko@mishhman·Oct 20

SpaceX delivered 2.7× the Street’s EPS estimate. Shortcut had already called for 1.7×. Shortcut is now 9–2 against Wall Street this week. $SPCX reported $0.35 EPS, above Shortcut’s $0.22 forecast and well above the Street’s $0.13 consensus. I bought three shares ahead of earnings based on that forecast. The Shortcut model forecast about $468m in Bitcoin gains, close to the $475m implied by the filings. Those gains explain roughly 17% of the EPS beat before tax effects.

Post attachment

I bought 3 shares of $SPCX, about $450, ahead of earnings. My Shortcut model predicts about $0.22 EPS versus the Street's $0.13, roughly 68.5% higher. It's the largest EPS gap in this week's forecasts. Part of the agent's thesis is that gains on SpaceX's roughly 18,700 Bitcoin could provide a bigger earnings boost than the Street's headline estimate suggests. The assumed gain contributes about $0.04 per share. I'm testing whether that forecast can translate into a profitable trade. My $APLD trade only made a few cents despite the model correctly predicting a large earnings beat. I exited at market close the following day, which didn't capture the surprise the way I'd hoped. Earnings accuracy and entry and exit timing are not trivial. Last week, Shortcut went 17–7 against Wall Street. This week, it's 5–1 so far. I'm forecasting 38 companies this week, including $TSLA, $SPCX, $META, $NFLX, and $IBM, using Opus 5.5 in Shortcut. I’m still a novice investor. Imagine what someone with years of experience could do with these tools. All 38 forecasts and financial models are public: beatthestreet.live/.

Post attachment
Richard Pham avatar
Richard Pham@phamrich_·Oct 16

the team has been hard at work not on traditional engineering but on architecting a system that allows us to ship 10x faster this is our first prod release since. check out everything we built in a week

New features on Shortcut shipped this week Review mode: You can now mark up your workbook and give Shortcut comments, like you would an analyst. Remote agent access: Connect Codex, Claude Code or your other favorite AI agents to Shortcut Cloud runs: Need to close your laptop? Hand off your task to a Shortcut cloud agent Recurring runs: schedule work for cloud agents to execute autonomously Full details: shortcut.ai/changelog/whats-…