MIT Researchers Built An AI Hedge Fund (Here's How To Use It) — Transcript

Full transcript

1. 0:00MIT and UCLA researchers have created
2. 0:02the very first true AI trading
3. 0:04framework. It's called Trading Agents.
4. 0:06It's run out of a lab called TA Research
5. 0:08and it basically acts as an AI hedge
6. 0:10fund. It's a team of AI agents,
7. 0:12analysts, researchers, traders, and a
8. 0:15risk manager with each one handling a
9. 0:17different part of the trading desk. So,
10. 0:19all you have to do is give a ticker or a
11. 0:21list of tickers from your portfolio or
12. 0:23watch list, and you'll see the entire
13. 0:25team of agents work through the company
14. 0:26and come back with research and
15. 0:28recommendations. So, in this video, I'll
16. 0:30show you how to actually use it and how
17. 0:32to set it up on Hostinger so that you
18. 0:34can always use it even when your
19. 0:35computer isn't on. And don't worry if
20. 0:38you have no coding background as I'll be
21. 0:39running through all the prompts step by
22. 0:41step to get it set up. Now, if it's your
23. 0:43first time on the channel, my name is
24. 0:44Brendan. I studied math and econ at
25. 0:46UCLA, worked in investment banking for 3
26. 0:49years, and have spent the past 2 years
27. 0:51building out trading systems for my
28. 0:52clients. And as always, this video is
29. 0:54not financial advice, and I'm not
30. 0:56guaranteeing any profits. So starting
31. 0:58off, this project is fully open source,
32. 1:00meaning that you can dive into the code,
33. 1:02make any changes you want, and more
34. 1:03importantly, it's completely free for
35. 1:05you to use. What I have here is the
36. 1:07actual GitHub repo. And if you're
37. 1:08unaware, GitHub is just where the code
38. 1:10for a project lives online. The
39. 1:12researchers behind Tar Research are
40. 1:14Yasao, Edward Sun, Dua, and Wei Wong.
41. 1:18And their affiliations are from UCLA and
42. 1:20MIT. So here inside the repo, you'll be
43. 1:22able to see how they designed the
44. 1:23project, which is Trading Agents, how
45. 1:25they tested it, and all the information
46. 1:27around code changes that they've made or
47. 1:29the source code as a whole. Now, because
48. 1:31Trading Agents is fully open source,
49. 1:33you'll be able to see all of the source
50. 1:35code here, as well as any new releases
51. 1:36they have. So you can see that version
52. 1:385.0 just released 4 days ago. Now,
53. 1:41anyone can download this repo, read
54. 1:42exactly how each agent is instructed
55. 1:44here, and then see how information from
56. 1:47one trading agent to another gets passed
57. 1:48on. One thing to be clear up front is
58. 1:50that even though the project is free and
59. 1:52open source, it doesn't mean that
60. 1:54running it is free. Because each of
61. 1:56these agents are going to be attached to
62. 1:57an AI model, those calls will cost money
63. 2:00per use. It can be from a few cents to a
64. 2:02dollar per run depending on the types of
65. 2:04AI models that you'll use. But later on
66. 2:06in the video, I'll show you the best
67. 2:07cost to performance ratio as to what you
68. 2:09should use inside this project. I'll
69. 2:11include the direct link to the GitHub
70. 2:13repo in the video description so that
71. 2:15you can go in, see the source code, and
72. 2:16run it yourself. And if you are
73. 2:18interested in learning how to run or
74. 2:20build off of more AI trading frameworks
75. 2:22or even building out your own AI trading
76. 2:24system, I've included a link in the
77. 2:26description to my school community. I
78. 2:28currently run the largest AI focused
79. 2:29trading community on school that covers
80. 2:31everything from stocks, options, crypto,
81. 2:33and futures. So, make sure to click the
82. 2:35link if you're interested. Now, before
83. 2:37we run anything, this is the actual
84. 2:38structure of trading agents. What we're
85. 2:40running here is basically an AI research
86. 2:42desk. So, we have 12 different agent
87. 2:44rules that are all connected to one
88. 2:46workflow. So basically each individual
89. 2:49agent here calls an assigned AI model
90. 2:51with its own instructions and just the
91. 2:53information it needs for its very
92. 2:55specific job. And if we were to dive
93. 2:57into every single agent, the market
94. 2:58analyst here just looks at price and
95. 3:00indicators. The fundamentals analyst
96. 3:02looks at financial statements, so
97. 3:04revenue, margins, debt. The news analyst
98. 3:06goes through the headlines. And then the
99. 3:08sentiment analyst looks at whatever
100. 3:10discussions it can pull up. So what
101. 3:11people are actually saying about a
102. 3:13certain company. The reports are then
103. 3:14fed into two researchers here. So the
104. 3:17bull researcher builds the positive case
105. 3:19basically why you'd buy something and
106. 3:21the bare researcher below pushes back
107. 3:23and looks for everything that could go
108. 3:25wrong. And then the research manager
109. 3:27here sits in between and weighs both
110. 3:29sides and then forms an actual
111. 3:30investment view. And then finally the
112. 3:32trader here takes that view and then
113. 3:34turns it into proposed plan. From that
114. 3:37proposed plan three different risk
115. 3:38reviewers look at the plan from three
116. 3:40different angles. So we have aggressive,
117. 3:43conservative and balanced. And then
118. 3:45finally, the portfolio manager here
119. 3:47reviews all of that and then produces
120. 3:48the final recommendation. So how this
121. 3:50differs from just chat GBT or Claude is
122. 3:53that if you were to ask them a bull and
123. 3:54bear case, you're going to get one
124. 3:56answer from that conversation. What
125. 3:58Taric research here does instead is that
126. 4:00it runs a defined chain of separate
127. 4:02agents and each one has a specific
128. 4:04responsibility, passes the work from
129. 4:07stage to stage and then saves every
130. 4:08report along the way. Now, one important
131. 4:10core concept here is that the models
132. 4:12that we largely discuss on the channel
133. 4:14are deterministic, meaning that they
134. 4:16have fixed rules and you're able to back
135. 4:18test them with this system here.
136. 4:20Instead, we have 12 different agents
137. 4:22that are all non-deterministic. This
138. 4:24means that their answers will slightly
139. 4:26vary each time you ask the same
140. 4:27question. And because of that, there's
141. 4:29no true way to actually back test this.
142. 4:31And this is why later on in the video,
143. 4:33I'll recommend you potentially use this
144. 4:35as part of research when looking at your
145. 4:37overall long-term portfolio rather than
146. 4:39something that just has strict rules for
147. 4:40intraday trading. Now, let's go over how
148. 4:42to actually run trading agents. So, as
149. 4:44you can see here, this is what the
150. 4:46interface looks like. We can see the
151. 4:48workflow steps are first the analyst
152. 4:49team, then the research team, then it
153. 4:51goes to the trader, the risk manager,
154. 4:53and then finally the portfolio manager
155. 4:55agent. And on step one, all you need to
156. 4:57do is type in the ticker. And then if
157. 4:59you need to, you can also include the
158. 5:00exchange suffix. But just to showcase,
159. 5:02I'll type in Nvidia. Press enter. Then
160. 5:06for step two, it's going to ask for the
161. 5:07analysis date. We'll just keep it as of
162. 5:09today, which is September 21st. So click
163. 5:11enter as well. Then we can select the
164. 5:13analyst team. So you would click space
165. 5:15on all the teams that you want to
166. 5:16include. So if you want a market
167. 5:18analyst, you would click space and then
168. 5:20you would go down. And then if you want
169. 5:21a sentiment analyst, you would click
170. 5:23space. You'd also click space for the
171. 5:24news analyst and the fundamentals
172. 5:26analyst if you want all of these agents
173. 5:28running. And then here we can see that
174. 5:29it's actually going through all of the
175. 5:31different steps. So all of the agents
176. 5:33are running. You can see the progress
177. 5:34between the analyst team and the
178. 5:36research team. You have the agents here.
179. 5:38You can see what's in progress, what's
180. 5:40in pending, and then all the messages
181. 5:42and tools as well. Now, while we wait
182. 5:43for this to run, the way you would use
183. 5:45this is that for example, if you were
184. 5:47interested in a ticker like Nvidia or
185. 5:49say Service Now, Coreweave, Micron, what
186. 5:52you would do is type in the tickers or
187. 5:54you can type in a list of tickers that
188. 5:55are in your portfolio. So, you do the
189. 5:57same thing for all of the different
190. 5:58tickers that you want to know more
191. 6:00about. And I treat this as research on
192. 6:02longerterm positions. And then one thing
193. 6:04to note is that each company goes
194. 6:06through its own research cycle. And
195. 6:08every run uses paid model calls. So if
196. 6:10you want to run this on six names,
197. 6:12you're paying for six different
198. 6:13analysis. Now, as you can see, it
199. 6:14finished the market analyst and it also
200. 6:16finished the sentiment analyst, news
201. 6:17analyst, and fundamentals analyst. So
202. 6:20what we saw was only one part of it.
203. 6:22Current report here is now inside the
204. 6:24research team. Um, it's currently gone
205. 6:26through around 6 minutes of analysis. So
206. 6:28again, this kind of strengthens why I
207. 6:30said you wouldn't use this for intraday
208. 6:31trading since it wouldn't be too helpful
209. 6:33to have this run for 10 minutes at a
210. 6:35time before deciding what to do during
211. 6:36the day. But this is pretty in-depth.
212. 6:39Each agent has its own tasks and I'll go
213. 6:41ahead and share all the reports once it
214. 6:43finally runs through everything. Now
215. 6:45finally after 8 minutes we can see that
216. 6:46the analysis is complete. So it ran
217. 6:49through all 12 agents. It used 17
218. 6:51different LLMs, 29 different tools. The
219. 6:54amount of tokens are here. The reports 7
220. 6:56out of seven and then the total runtime
221. 6:58was 8 minutes. So like I explained
222. 7:00during the structure, this portfolio
223. 7:02manager agent is the last one that reads
224. 7:04through everything and the final rating
225. 7:06was hold. I'll go through some of the
226. 7:08other sub aents, but some set to buy,
227. 7:09some set to hold, and then this
228. 7:11portfolio manager agent is the one that
229. 7:13makes the final decision. This tells you
230. 7:15to maintain Nvidia at a neutral
231. 7:16portfolio weight. You have the
232. 7:18investment thesis as to why it's saying
233. 7:20this. So, this includes all the macro
234. 7:22market analysis, individual technical
235. 7:24analysis, as well as fundamentals about
236. 7:26the company. So, now what we want to do
237. 7:28is save this report. And this is the
238. 7:30final report. So, this is what every
239. 7:31individual agent went through. We first
240. 7:33have the company market profile. This is
241. 7:35just all the basic stats that we need.
242. 7:37Then we have the latest quarterly income
243. 7:39statement. This is just needed for the
244. 7:40fundamental analyst. So let's just go
245. 7:42ahead and scroll through all these
246. 7:43fundamentals since this is the data in
247. 7:46to form either the bull or bear case
248. 7:48that I explained. So if you scroll all
249. 7:49the way down into section 8, we have the
250. 7:51bull case of Nvidia. We have the bare
251. 7:54case of Nvidia. So these are all the
252. 7:56different components that go into what
253. 7:58the next agent thinks about. So here we
254. 8:00can see that the initial recommendation
255. 8:02was to buy based on the bare and bull
256. 8:04case of fundamentals. This is all the
257. 8:06interpretations of Nvidia. But then we
258. 8:08move on to the research team. This again
259. 8:10is an entirely separate subset of
260. 8:12agentic teams that have a bull
261. 8:14researcher. Then if we scroll down we
262. 8:16have a bear researcher as well. So when
263. 8:18you run this yourself feel free to read
264. 8:19through all of these reports since this
265. 8:21is important for you to decide what to
266. 8:23ultimately do. Um I'm not going to do
267. 8:25that cuz I don't want this to be a
268. 8:262-hour long video. So I'll just scroll
269. 8:28through the bear researcher as well, but
270. 8:30it's important to know that all of these
271. 8:31just feed into the next step. So while
272. 8:33we have the fundamental analyst saying
273. 8:35buy, this research manager is saying
274. 8:37hold, and then we have a trader that has
275. 8:39a completely separate decision of hold
276. 8:41as well. All of this goes all the way
277. 8:44down into the portfolio manager with the
278. 8:46final recommendation. This way you don't
279. 8:48just see this final output of hold, you
280. 8:50see all the different components that
281. 8:52flow into it. see if you agree or
282. 8:54disagree with each one, but each one
283. 8:55does have its own intelligent process
284. 8:57and thinking, meaning that none of, for
285. 8:59example, the fundamental analyst will
286. 9:01ever impact or affect what the market
287. 9:03analyst thinks and nothing that the
288. 9:05market analyst thinks will ever impact
289. 9:06or affect what the next agent thinks.
290. 9:09And the way I would think about this
291. 9:10again is to run this for fairly
292. 9:12long-term positioning. Here we can even
293. 9:14see the time horizon is to hold for 3 to
294. 9:166 months. If you want to, you can rerun
295. 9:18the same analysis every single day
296. 9:20depending on how many tokens you want to
297. 9:22spend or every single month if you have
298. 9:23a more long-term portfolio. Now, if you
299. 9:26want to do multiple tickers, so for
300. 9:28example, if you have a watch list or
301. 9:29your entire portfolio, you can use this
302. 9:32tool called watch list research from
303. 9:34Trading Agents, it's the same exact
304. 9:35thing as what I just ran through, except
305. 9:37now you can input multiple tickers. And
306. 9:40because I'm running this through
307. 9:41Hostinger, which I'll show you how to
308. 9:42set up later, all the reports will be
309. 9:44saved on this VPS. So, we have a save
310. 9:46decision journal and then once you want
311. 9:48to input more tickers, it'll all be
312. 9:50saved inside the same VPS. And even
313. 9:52though there's no inherent back testing
314. 9:54with these non-deterministic systems,
315. 9:56this trading agents tool from the TA
316. 9:58research team does have the ability to
317. 10:00go back in time and see what this model
318. 10:02would have said about a ticker, say a
319. 10:05month ago, a year ago, or even further
320. 10:07out. So, in step two, after you input
321. 10:09the ticker, you can input the date. So,
322. 10:11I would say this is the closest thing
323. 10:12you could get theoretically to a back
324. 10:14test. Now, one last thing before I show
325. 10:16you how to set this up yourself is that
326. 10:18this shows the different costs as well
327. 10:20as models you should use for the
328. 10:22structure. In that initial run I just
329. 10:23did, I used GPT 5.6 soul for everything.
330. 10:27But, as you can imagine, it gets pretty
331. 10:28costly. Down here, there's a chart
332. 10:30showing the cost. So, per 1 million
333. 10:32tokens, it's $4 and the output is $20.
334. 10:35For this run, we used around 100,000
335. 10:37tokens. So, the total cost was $1.26.
336. 10:40Now, obviously, if you have a watch list
337. 10:42and you're running this every single
338. 10:43day, it's going to cost you a few
339. 10:44dollars every single day. But that's why
340. 10:46I would recommend something like Kimmy
341. 10:48K2.6, this is significantly cheaper than
342. 10:51GPT 5.6 Soul at only 95 and $4 per 1
343. 10:55million tokens, meaning that it's around
344. 10:5780% cheaper. So, something that could
345. 10:59cost a dollar per run would be closer to
346. 11:0110 or 20. Now the performance between
347. 11:03the two are very similar. But if you do
348. 11:06want to, for example, have the strongest
349. 11:08models on certain managers. I
350. 11:10highlighted these orange as deep
351. 11:12managers. So the portfolio manager and
352. 11:14research manager, these are the two
353. 11:16agents that actually take in most of the
354. 11:18information and make the most thorough
355. 11:19outputs as well. And because of this,
356. 11:21these are the two managers that I would
357. 11:23recommend having the strongest possible
358. 11:25models for. So an example of how this
359. 11:27could look like is having Kimmy on every
360. 11:29single model here and then switching
361. 11:30over to 5.6 6 soul as the research
362. 11:32manager and then again Kimmy on all of
363. 11:34these and then once again switching back
364. 11:36to 5.6 soul for the portfolio manager
365. 11:39and when I set this up in chat GBT or
366. 11:41claude you'll be able to see that you
367. 11:43can configure every single model
368. 11:44differently for each agent if you do
369. 11:46want this to run all the time without
370. 11:48your computer needing to be on. You can
371. 11:50use a VPS like Hostinger. So I'll go
372. 11:51ahead and show you the setup through
373. 11:53creating your own server which plans to
374. 11:55sign up for and then connecting
375. 11:56everything together. Now, what I would
376. 11:58do is first go over to the link
377. 11:59hostinger.com/ipipathways10,
378. 12:02which will also be in the description.
379. 12:04And then once you're on this page, you
380. 12:06can go ahead and choose plans for your
381. 12:07virtual private server. And for what I
382. 12:09used in the beginning of the video for
383. 12:10the demo, I used KVM2 since it's the
384. 12:12perfect mix of CPU and RAM usage. But if
385. 12:15you do see yourself needing either more
386. 12:17RAM or CPU, you can go ahead and
387. 12:18upgrade. But for our choices here, we
388. 12:20can even start with one or two depending
389. 12:22on how many watch list or workflows you
390. 12:24want to enable with this. Now, the cost
391. 12:26is around $10 per month to have your own
392. 12:27VPS, which again will always stay on
393. 12:30versus running something directly on
394. 12:31your computer. And if you haven't
395. 12:33already, I recommend watching my Hermes
396. 12:35video that I made a few months back
397. 12:36since it goes over the entire setup
398. 12:38similar to this. So this way, your
399. 12:40Hostinger VPS isn't just used for this
400. 12:42one specific task. You can have your
401. 12:44server run Hermes, Trading Agents, or
402. 12:46any other projects that you want
403. 12:47directly on the cloud versus on your
404. 12:49local computer. And if you do want a
405. 12:51coupon code just to save a bit of money
406. 12:52on the 12 or 24month plans, you can just
407. 12:55type in AI Pathways here and click
408. 12:57apply. This way you'll get 10% off the
409. 12:58cost to have your VPS up and running.
410. 13:00Now, once everything looks good, you can
411. 13:02just go ahead and click continue. Now,
412. 13:03after you choose your plan, it should
413. 13:05prompt you to go through the one-time
414. 13:06setup wizard where you would first
415. 13:08choose a location for your server. After
416. 13:10all that, you should be taken to this
417. 13:12page where you can see your server
418. 13:13successfully running. So, you have your
419. 13:15IP address, the status of the server, as
420. 13:17well as when it expires. Then all you'll
421. 13:19need to do is go ahead and to click
422. 13:21manage and then go to the web console.
423. 13:23It'll then open up this console where
424. 13:24you can type in terminal commands to
425. 13:26download Python to be able to run the
426. 13:28repo. This is also where you would
427. 13:30download all the contents of that repo
428. 13:32and be able to run it directly inside
429. 13:33this window here. Now the next step
430. 13:35could either be super simple or you can
431. 13:37type in the commands yourself. Since I
432. 13:39want to make this as simple as possible
433. 13:40for people on the channel, we can go
434. 13:42ahead and literally ask Chat GBT codeex
435. 13:45to do this for us. If you're on any sort
436. 13:47of paid plan through chatbt, it can go
437. 13:49ahead and access Chrome, which is your
438. 13:51browser, and be able to run all the
439. 13:52commands for you. But if you don't have
440. 13:54a Chat GBT or clawed account, then you
441. 13:56can go ahead and just manually type in
442. 13:58the commands here as well. I'll go ahead
443. 14:00and flash them on the screen. But just
444. 14:01to show you how simple this can be, let
445. 14:03me go ahead and open Chad GBT codeex so
446. 14:05you can see all the commands it's going
447. 14:06to run. Now, the next step is super
448. 14:08simple. Make sure to open Chad GBT
449. 14:10codeex on your desktop. This is what
450. 14:12allows Chad GBT to control your
451. 14:14computer, look through your browsers,
452. 14:16and in our case, look through our Chrome
453. 14:17browser, which is where we set up the
454. 14:19HostNerVPS. Then all we need to do is
455. 14:21say that our hostingervp is ready in
456. 14:23Chrome. We want to install trading
457. 14:24agents from and then the GitHub link on
458. 14:27it. We want to use Kimmy K2.6 as our
459. 14:29model. So it will stop us before the API
460. 14:32key selection part where we can type in
461. 14:34our own API keys securely. And then
462. 14:36finally, we want to keep the original
463. 14:38project code unchanged and to stop
464. 14:40before ever running any analysis. So all
465. 14:42you have to do is click enter here and
466. 14:44it'll actually go into your Chrome and
467. 14:46set up all the commands for you. So here
468. 14:48we can see that chat GBT is now running
469. 14:50through all the commands needed. So I'm
470. 14:52not touching anything. We can see that
471. 14:54chat gbt started debugging on this
472. 14:56browser. It's running its own commands
473. 14:57to be able to install the repo and get
474. 14:59it ready for us. And then again, if you
475. 15:01don't want to use chat GBT, just run the
476. 15:03same exact commands that it did here,
477. 15:05and you'll be able to use it as well. So
478. 15:07next, it's gone ahead and installed
479. 15:08everything for us. And this is where we
480. 15:10can securely type in our API keys. So
481. 15:12when prompting with Chad GBT, we told
482. 15:14them to use Kimmy K 2.6. But obviously,
483. 15:17if you want to change this to OpenAI,
484. 15:19just type in your OpenAI key here and
485. 15:21then have Chad GBT or you can do this
486. 15:23yourself change this to the exact chat
487. 15:25GBT model. And the easiest way to
488. 15:27actually use that Kimmy 2.6 6 API key is
489. 15:30to get it from open router which is the
490. 15:32website here. You'll be able to access a
491. 15:34bunch of different API keys. So just
492. 15:36choose the one that you feel fits best.
493. 15:38In my opinion, Kimmy is probably the
494. 15:40cheapest value option that you can
495. 15:42reasonably use inside this model. And
496. 15:44then down here below, you can actually
497. 15:46set Fred API keys as well. So this is if
498. 15:49you need macro data. So that's one of
499. 15:51the agents that I showed in the
500. 15:52beginning of the video that actually
501. 15:53looks at macro sentiment and macro data.
502. 15:56So, if you want to have that, just go
503. 15:57ahead to this link right here,
504. 15:59fred.saint louisfed.org. So, this API
505. 16:02key is completely free. You just need to
506. 16:04make an account here as well. And then
507. 16:05once you submitted all your API keys,
508. 16:07you're pretty much good to go to start
509. 16:08running this yourself. Then once you
510. 16:10finish inputting the API keys, you can
511. 16:11let Chad GBT know that it's ready and
512. 16:13it'll go ahead and run this trading
513. 16:15agents for you. Now, this is again for
514. 16:17Hostinger if you want your own VPS and
515. 16:19you can combine it with something like
516. 16:20Hermes or another scheduler to make sure
517. 16:23that this runs every single day and
518. 16:24input it into your own agentic
519. 16:26workflows. But if you want this just to
520. 16:28be local on your computer without a VPS,
521. 16:30you can do the same exact steps. The
522. 16:32only thing that's different is that you
523. 16:33wouldn't need to configure Hostinger as
524. 16:35a VPS. You can just save everything
525. 16:37directly to your computer. So, open up
526. 16:39VS Code, use Claude Code, or use the
527. 16:41Codeex app once again. Tell them to
528. 16:43install and run everything for you. but
529. 16:45this time again not through Hostinger.
530. 16:47You would just download it all directly
531. 16:48on your computer. Now, if you found this
532. 16:50video helpful, make sure to leave a
533. 16:51like, comment, and subscribe as it
534. 16:53greatly helps out the channel. And like
535. 16:55I mentioned in the beginning of the
536. 16:56video, if you do want to dive deeper
537. 16:57into using AI for trading as well as
538. 16:59investing, make sure to click the link
539. 17:01in the description.