WEBVTT

NOTE Sentence-level transcript of https://www.youtube.com/watch?v=ZMZlAn6Pyfg

NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/ZMZlAn6Pyfg.html. A cue ends where the next begins, or 2 s after its last word.

s1
00:00:01.309 --> 00:00:03.309
[music]

s2
00:00:12.400 --> 00:00:13.520
My name is Harry.

s3
00:00:13.520 --> 00:00:19.360
Um, I am the head of AI at Contour Asset Management, which is a hedge fund based in the city.

s4
00:00:19.360 --> 00:00:25.320
It's been around for a little over a decade now, um, but I've been here for the past year-ish.

s5
00:00:25.320 --> 00:00:32.200
Um, we're a fundamental long short shop, um, around like 15 20 people, 6 billion AUM.

s6
00:00:32.200 --> 00:00:33.405
Um, and mostly

s7
00:00:33.405 --> 00:00:33.560
[snorts]

s8
00:00:33.560 --> 00:00:36.080
invest in the TMT space.

s9
00:00:36.080 --> 00:00:50.040
Um, so a lot of our investors are pretty tech forward by nature, but want to bring a little bit more, um, AI workflow into the investment firm itself and see, okay, we've had all these processes done

s10
00:00:50.040 --> 00:00:52.680
for many years in a very similar way.

s11
00:00:52.680 --> 00:01:03.000
How can we kind of shake things up and use AI to our advantage, especially when it comes to the research process, but then also bring that into trading, operations,

s12
00:01:03.000 --> 00:01:06.320
compliance, finance, pretty much all parts of the firm.

s13
00:01:06.320 --> 00:01:18.120
So, came in in the past year and basically my mandate was, how can I show impact across all areas of the firm, largely starting with this research process, right?

s14
00:01:18.120 --> 00:01:31.360
And as probably all of you guys in this room can understand and appreciate, it's a really big challenge to show, hey, we're spending this amount of money and time on AI spend and impact.

s15
00:01:31.360 --> 00:01:33.880
What are we actually gaining from it?

s16
00:01:33.880 --> 00:01:45.400
Both from a P&L standpoint, which is to come, um, but also just from a more operational and are we getting good ideas perspective, too.

s17
00:01:45.400 --> 00:01:57.960
The ROI capture from our investment is a really big challenge to understand and capture, but I want to talk a little bit about some of the ways at in the investing landscape

s18
00:01:57.960 --> 00:02:05.560
we're able to start to think about hey, we've put in this amount of money into our AI spend, what are we getting out of it?

s19
00:02:08.240 --> 00:02:09.440
Cool.

s20
00:02:09.440 --> 00:02:20.840
So, a few things that we are thinking about here and intentionally trying to keep this a little bit vague and more in terms of the operational and process parts of it, but down to dive into any technical stuff

s21
00:02:20.840 --> 00:02:25.280
either at the end of the session or afterwards as well.

s22
00:02:25.280 --> 00:02:33.200
So, what we're trying to do is basically bring out live AB tests for some of the research workflows that our analysts have.

s23
00:02:33.200 --> 00:02:37.880
So, we have around 10 analysts that each cover different parts of our coverage, right?

s24
00:02:37.880 --> 00:02:46.320
Which is how it works at a lot of these um hedge funds, especially on the smaller side, where you have maybe one person that covers 30 to 40 names.

s25
00:02:46.320 --> 00:03:01.760
And they've had their uh existing processes in terms of how they think about their investment thesis, how they actually come up with new ideas, and then validate and test out over the long term whether those ideas actually came to fruition.

s26
00:03:01.760 --> 00:03:15.400
But now in a AI and more agentic world when it comes to investing, the challenge is would you have come to the same type of conclusion with the AI tools or without the AI tools as well?

s27
00:03:15.400 --> 00:03:30.680
And a process that we've done in the past few months that I think has been pretty successful has been actually creating live AB tests by recording and documenting their entire process that they would have done in a pre-AI world

s28
00:03:30.680 --> 00:03:35.760
and how they're doing it as we roll out more workflows and tools in the company itself.

s29
00:03:35.760 --> 00:03:43.160
So, for example, what we did is for one specific name when it came to our three-month cycle.

s30
00:03:43.160 --> 00:03:50.560
So, from one earnings report to the following one, we tracked if they didn't use any AI tools.

s31
00:03:50.560 --> 00:03:56.600
And this is at first just hey, if you're not using any LLMs, what does your workflow look like?

s32
00:03:56.600 --> 00:04:08.040
But then, as we've introduced more um tools at the hedge fund itself, like okay, you have some tools that help you do expert call interviews better, or ones that help you field a consumer survey

s33
00:04:08.040 --> 00:04:11.080
that's done by an agent instead of a human.

s34
00:04:11.080 --> 00:04:14.160
What are your takeaways without things like that?

s35
00:04:14.160 --> 00:04:22.560
And then we're documenting, okay, you've done this process without AI, but also without these additional tech workflows that we're building out.

s36
00:04:22.560 --> 00:04:25.040
What are the data sources that you're finding?

s37
00:04:25.040 --> 00:04:28.560
And what are the conclusions you're coming out with?

s38
00:04:28.560 --> 00:04:33.720
Then we run that same process involving our PM including some of these AI tools.

s39
00:04:33.720 --> 00:04:37.600
And we're building them one by one with different things that we're coming at with, right?

s40
00:04:37.600 --> 00:04:50.040
So, for example, when it comes to expert call interviews, one thing that's been pretty helpful is you have expert calls that are run with an agent, and then you synthesize the notes yourself.

s41
00:04:50.040 --> 00:05:00.800
How does that process work when you do it with an agent work versus if you have the agent synthesize the notes for you, and then you come up with your conclusions at the very end of it, right?

s42
00:05:00.800 --> 00:05:11.720
So, it's kind of thinking about the question of how long would it have taken to source some of this information manually versus if you were able to do it in a more agentic manner.

s43
00:05:11.720 --> 00:05:22.040
This This is definitely a big challenge when it comes to the investing landscape because a lot of the work that we're doing is just based off of idea generation.

s44
00:05:22.040 --> 00:05:33.320
So, you're coming up with the new ideas and new theses, and you want to see, okay, can they be proven out in a short-term scale, which is a little less of what we do, but at least in like the 3-month window,

s45
00:05:33.320 --> 00:05:42.560
or are they things that take 6 months or a year to manifest in terms of an actual um impact to stock price, right?

s46
00:05:42.560 --> 00:05:48.920
So, the challenge with something like this is you can't show ROI immediately.

s47
00:05:48.920 --> 00:05:52.160
It's more so from a process-based manner, right?

s48
00:05:52.160 --> 00:05:57.160
So, are we improving the process of how we're conducting consumer surveys?

s49
00:05:57.160 --> 00:06:00.240
Are we improving the process of how we're updating our models?

s50
00:06:00.240 --> 00:06:13.440
And one by one, we're trying to bring out a little bit more of um uh ROI measurement when it comes to okay, are we actually changing the way someone's working

s51
00:06:13.440 --> 00:06:17.680
or we just adding more thrust to their process?

s52
00:06:19.160 --> 00:06:24.120
So, talked a little bit about these use cases, but maybe just to touch a little bit more on them.

s53
00:06:24.120 --> 00:06:35.960
The main goal when it comes to um fundamental investing is thinking of a thesis, but also as I've learned, it's to not be so anchored to a thesis that you come up with initially

s54
00:06:35.960 --> 00:06:45.960
and actually trusting some of the data, but also the expert and other insights that you're getting to remove some of those anchors as well, right?

s55
00:06:45.960 --> 00:06:52.200
And this is where I think that there's a lot of impact measurement when it comes to AI tooling that can be built.

s56
00:06:52.200 --> 00:07:11.440
So, example, um we have manual idea generation agents that are being run to take different ideas that are being come up with per name per analyst and adding more layers in terms of okay, you've come up with a few different data points from different alternative data sets

s57
00:07:11.440 --> 00:07:15.760
or you've talked to an expert in the industry to formulate some of that idea.

s58
00:07:15.760 --> 00:07:28.760
Let's layer on a few more conclusions or parts of the process to think about when it comes to how are you actually how are you actually understanding what that idea is going to materialize

s59
00:07:28.760 --> 00:07:35.520
in the next few months or if it's something that it takes even longer to take to take into account into the market.

s60
00:07:35.520 --> 00:07:47.000
Additionally, we're thinking a little bit more as to one of the biggest problems that we're seeing is you take an idea or you have an idea, you might work on it for a couple of days,

s61
00:07:47.000 --> 00:07:53.240
and then that idea might not be interesting to the PM or something else takes precedence, and you lose that idea.

s62
00:07:53.240 --> 00:07:57.720
And it doesn't come back for months, and then maybe you miss like a key price window.

s63
00:07:57.720 --> 00:08:13.560
So, having agents that are also tracking through all of the notes that you've been taking, even if you only worked on an idea for a day, tracking it against the market and trying to see, "Okay, is this now an idea that's interesting, and should we flash that as a point to not just you, but the PM and

s64
00:08:13.560 --> 00:08:14.880
the trader as well?"

s65
00:08:14.880 --> 00:08:23.480
And bringing out more quick work when it comes to you've had a bunch of different things you've worked on, and maybe it's been weeks since you've looked at it.

s66
00:08:23.480 --> 00:08:27.360
Now, let's get back to it because it's interesting again, right?

s67
00:08:27.360 --> 00:08:34.880
I think that's a big problem that a lot of hedge funds have had, where you have a bunch of different things that you can be working on, you have to prioritize,

s68
00:08:34.880 --> 00:08:42.120
but the timing is actually the most important thing as to when you want to strike when it comes to an interesting point in the market, right?

s69
00:08:42.120 --> 00:08:56.920
So, building agents that are not only doing idea a generation, but thinking about timing and when you should enter exit a position, even if it's not something that has been top of mind for you, is something that we've been really thinking about as well.

s70
00:08:56.920 --> 00:09:06.400
Um and I guess the last point that I want to make here, thinking a little bit more about the alternative data options when it comes to investing workflows, right?

s71
00:09:06.400 --> 00:09:13.240
A lot of funds spend quite a good sum of money when it comes to acquiring really new and novel data sets, right?

s72
00:09:13.240 --> 00:09:25.160
And I think in this age, there's a bunch of different providers that are coming out with very nuanced solutions, even when it comes to things like satellite mapping or really granular credit card data.

s73
00:09:25.160 --> 00:09:29.720
Those things are helpful, but they're only helpful when they're all brought together, right?

s74
00:09:29.720 --> 00:09:40.640
So, if you think about something like a Yipit, which provides credit card panel data, that lives in isolation, but now with MTP and API access for things like Gipit,

s75
00:09:40.640 --> 00:09:43.240
you can feed it into our agents, which is kind of what we've done.

s76
00:09:43.240 --> 00:09:53.720
So, it takes into account things like your expert call or notes transcripts, but also layers in different data sets when it comes to credit card panels or spend data

s77
00:09:53.720 --> 00:09:55.640
to see how those two things interact.

s78
00:09:55.640 --> 00:10:01.800
Where before those have been differentiated, now we're trying to curb that issue as well.

s79
00:10:05.000 --> 00:10:12.160
Additionally, I think a challenge with measuring ROI is the time aspect of it.

s80
00:10:12.160 --> 00:10:22.040
So, you spend a lot of money when it comes to acquiring some of this tooling when it comes to research, model updates, AI spend.

s81
00:10:22.040 --> 00:10:29.080
But, there's a lot of time when it comes to what is an analyst spending their day on, which is the main thing that we're thinking about, right?

s82
00:10:29.080 --> 00:10:40.160
So, what we've also done here, which I think has been pretty useful, is develop metrics when it comes to hey, these are the insights that have been generated by an analyst per day.

s83
00:10:40.160 --> 00:10:41.920
What was the time spent on that?

s84
00:10:41.920 --> 00:10:46.920
So, tracking time spent on each name.

s85
00:10:46.920 --> 00:10:52.000
And what were the eventual P&L outcomes from some from a decision like that?

s86
00:10:52.000 --> 00:11:00.280
That is something that is good in investing landscape because you have a very very obvious success metric when it comes to P&L.

s87
00:11:00.280 --> 00:11:13.080
But, it takes a lot longer than other industries to actually materialize, especially from a more fundamental shop, where you're not trading by the day, but you're making a few trades a day that are expected to materialize

s88
00:11:13.080 --> 00:11:16.480
over months-long horizons, right?

s89
00:11:16.480 --> 00:11:21.360
So, understanding where analyst time goes is really valuable.

s90
00:11:21.360 --> 00:11:33.920
And tracking things like a cost per useful insight, which is the metric that we've come up with, really helps us understand, okay, you spent your time to generate these ideas, and those are all being tracked per day.

s91
00:11:33.920 --> 00:11:35.880
Where do those ideas eventually land?

s92
00:11:35.880 --> 00:11:37.600
What were the interesting ones?

s93
00:11:37.600 --> 00:11:43.400
And what were the tools that were taken to come up with an idea like that?

s94
00:11:44.760 --> 00:11:56.160
So, that largely touches on the research side of workflows when it comes to measuring um our ROI from not just AI spend, but general tech investment.

s95
00:11:56.160 --> 00:12:00.360
But, we want to also improve this across the entire firm, right?

s96
00:12:00.360 --> 00:12:03.040
It's really important to make sure that we do quality research.

s97
00:12:03.040 --> 00:12:04.840
That is how you get ahead.

s98
00:12:04.840 --> 00:12:15.560
But, if you have better and more involved trading, as well as stronger ops and finance, that just helps the firm at large to free up time to improve processes, right?

s99
00:12:15.560 --> 00:12:22.320
So, for example, on the trading side of things, we're doing a lot more when it comes to measuring our entry and exit times.

s100
00:12:22.320 --> 00:12:28.400
Thinking about how our position is spread across multiple different industries and multiple names within an industry.

s101
00:12:28.400 --> 00:12:37.360
And having AI agents track those names to make sure that our traders are on top of it and understand, okay, these are interesting insights.

s102
00:12:37.360 --> 00:12:44.839
So, bring them much more into the analyst world because their time should not just be spent on executing orders.

s103
00:12:44.839 --> 00:12:51.960
It should also be spent on bringing idea generation, feeding ideas to our analysts and associates as well.

s104
00:12:51.960 --> 00:13:02.680
Additionally, on the op side, this is something that I think a lot of um general companies can appreciate, but it's been pretty interesting to see it from a hedge fund landscape.

s105
00:13:02.680 --> 00:13:12.520
Many of the workflows that are being done are basically set within Bloomberg or set within your um order keeping agent.

s106
00:13:12.520 --> 00:13:22.920
And those, while they've come up with some cool AI tools to help improve the process, it's really challenging to integrate those together, which is something that we're starting to do here.

s107
00:13:22.920 --> 00:13:35.680
So, how can you take all parts of the ops and compliance workflow and marry them into one one specific tool that our ops and compliance um functions can use.

s108
00:13:35.680 --> 00:13:48.160
So, they're not just reliant on external providers to give them piecemeal solutions to hey, this is how you can do order reconciliation, or this is how you can uh track vendor info.

s109
00:13:48.160 --> 00:13:52.200
The goal that we're trying to create here is obviously minimize repetitive work.

s110
00:13:52.200 --> 00:13:54.280
That's what everyone here is thinking about.

s111
00:13:54.280 --> 00:14:01.040
But, specifically from financial operations, that is something that can really be consolidated into one platform.

s112
00:14:01.040 --> 00:14:07.200
Where right now you have to kind of go into many different platforms for different pieces of the order flow.

s113
00:14:07.200 --> 00:14:12.080
So, bring those all into one is um is a pretty big priority for us.

s114
00:14:12.080 --> 00:14:23.400
And I think not necessarily to the point of this talk, a way to really capture ROI, but to understand how can you take people who are working in trading and ops right now,

s115
00:14:23.400 --> 00:14:34.160
and use their skills to improve the main goal of the fund, which should be to create create new ideas, understand the market, and focus on timing.

s116
00:14:35.880 --> 00:14:36.440
Okay.

s117
00:14:36.440 --> 00:14:47.160
The the last piece that I want to talk about here is thinking a little bit about the existing spend that we have um when it comes to different data sources that we're working with,

s118
00:14:47.160 --> 00:14:52.680
and how we can kind of up-level that when it comes to hedge fund workflows.

s119
00:14:52.680 --> 00:15:06.760
So, right now a lot of hedge funds will spend on either they're spending on different piecemeal solutions like an AlphaSense for for data, a Yipit for alt data, or they're spending with um

s120
00:15:06.760 --> 00:15:11.400
uh holistic solutions that use um third party.

s121
00:15:11.400 --> 00:15:16.440
So, something like buying a Roqoo membership, or investing in like Kensho, right?

s122
00:15:16.440 --> 00:15:28.440
Those are really great solutions, and I think the the goal of this eventually that we're trying to build in house, is how can you kind of bring all these tools that different analysts use at a hedge fund

s123
00:15:28.440 --> 00:15:34.600
into one platform, but also up level the spend that we have right now.

s124
00:15:34.600 --> 00:15:45.600
So, for example, with an existing, let's say, AlphaSense subscription, which is a big um model and call plus data provider.

s125
00:15:45.880 --> 00:15:53.240
Before, a lot of analysts will just use the AlphaSense portal and pull from there and work within their existing models.

s126
00:15:53.240 --> 00:16:04.000
Things like Rogo will kind of take that one step further by allowing you to have all that data within your model itself, and you can have a platform that you chat with Rogo agents for.

s127
00:16:04.000 --> 00:16:18.080
What we're trying to do is still use things like an AlphaSense or Yipit tool, but bring them into our own platform internally that learns from our analysts analyst habits and is programmed specifically for each individual analyst

s128
00:16:18.080 --> 00:16:21.920
and their model pattern, so it's not all as one.

s129
00:16:21.920 --> 00:16:24.920
And how can you increase the value of your subscription?

s130
00:16:24.920 --> 00:16:35.920
A lot of these companies, obviously, are having like MCPs and APIs that are being pulled out, but the goal should be to learn continuously from all of the different types of data

s131
00:16:35.920 --> 00:16:40.000
and the model updates that we're getting from this as well.

s132
00:16:40.600 --> 00:16:50.400
And then, to wrap up here, I think the the last point of the ROI equation, too, has been a pretty interesting one on my end when it comes to, "Okay,

s133
00:16:50.400 --> 00:17:01.520
you're bringing in new tooling to people who are very tech forward for investors, but you understand that there are going to be places where this breaks down, where model calls break or the connector

s134
00:17:01.520 --> 00:17:05.000
breaks down every couple days, and then it's your job to fix it, right?"

s135
00:17:05.000 --> 00:17:20.800
So, part of the ROI equation that we've been taking into account at Contour is trying to understand, "Okay, there are these pauses that we're getting, but also what is the time that that being taken away trying to fix bugs, that someone from the tech side cannot fix itself, right?

s136
00:17:20.800 --> 00:17:27.560
So, that's something that I think is pretty important to consider when it comes to, "Hey, we're involving AI in investing workflows."

s137
00:17:27.560 --> 00:17:37.040
It's all really exciting, and people are very into the concept of it, but we also need to understand, okay, when it comes to time spent for, let's say, valuable insight,

s138
00:17:37.040 --> 00:17:41.200
what is the time detracted to understand how these tools work as well.

s139
00:17:41.200 --> 00:17:47.360
And that's something where that time is decreasing as we go forward, but it's a metric that we're still considering.

s140
00:17:47.360 --> 00:18:03.560
Um so, I think to wrap up, in investing workflows, measuring ROIs definitely a huge challenge for us, but it's really important to be involved in this space because as AI adoption increases, especially at a lot of these smaller fundamental shops,

s141
00:18:03.560 --> 00:18:10.920
having one unified layer that's doing a lot of different tasks across analysts is going to be really, really valuable.

s142
00:18:10.920 --> 00:18:16.080
Um I think that's all the time I have, but if anyone has any questions after the talk, would love to connect.

s143
00:18:16.080 --> 00:18:18.280
Thank you.