Claude vs ChatGPT: Who Can Build a Profitable Trading Strategy? — Transcript

Full transcript

1. 0:00Hello guys and welcome back to the
2. 0:01channel. Today we're going to find out
3. 0:03if two of the most powerful AI models
4. 0:05right now can actually build profitable
5. 0:08trading strategy. Opus 5.5 versus GPT
6. 0:11salt 6.1.
7. 0:13Both going to have the same lab, the
8. 0:15same data set, same asset and one single
9. 0:18idea to start from and at the end only
10. 0:21one thing matters. Which one walks out
11. 0:23with a strategy that actually makes
12. 0:25money. But before we get into it, make
13. 0:27sure to check out the current discount
14. 0:29bonus that's running. They offer
15. 0:31options, the classic evaluation, a
16. 0:33one-time fee evaluation with no
17. 0:35activation fee after passing and a
18. 0:37straight to funded option. You can use
19. 0:40the coupon code stranger for the best
20. 0:42available discount. So spoiler alert,
21. 0:45one of them actually did it. It built a
22. 0:47profitable strategy that survived
23. 0:49everything I threw at it. If you want a
24. 0:51PDF with the full rules and the logic
25. 0:53behind it, all you have to do is
26. 0:56subscribe and leave a comment down below
27. 0:58and [snorts] I'll send it to you. Now,
28. 1:00how does this work? Both models will get
29. 1:03the exact same lab, the research
30. 1:05infrastructure that I built and use
31. 1:07every single day. It's called strategy
32. 1:09architect.
33. 1:10Two identical copies with a wall between
34. 1:13them so no picking and no copying each
35. 1:16other. Inside they will get eight years
36. 1:18of data to learn from from 2016 to 2023
37. 1:22and they can test it, learn from it,
38. 1:24break it, rebuild it, whatever they
39. 1:26want. At the end each one hands me its
40. 1:29best strategies. We're going to call
41. 1:30them candidates. But anything can look
42. 1:33good on data that they have already
43. 1:35studied. So I locked away two and a half
44. 1:37years from 2024 to August 2026 that
45. 1:42neither of them has ever seen. That's
46. 1:44the final exam and that's where we find
47. 1:46out if their candidates actually work. I
48. 1:49also didn't want them starting from a
49. 1:51blank page with 20 assets
50. 1:55>> Cuz then you can not really compare
51. 1:56anything. So, I gave them one asset and
52. 1:59one idea. The asset is US dollar against
53. 2:02the Japanese yen, one of the most traded
54. 2:04pairs in the world, super liquids with
55. 2:06tight spreads, and the idea is this.
56. 2:08USD/JPY often builds a range during the
57. 2:11Asian session, and London tends to break
58. 2:14it. So, find me a profitable trading
59. 2:16strategy based on that.
60. 2:19So, as you can see, these are the labs.
61. 2:20This one belongs to Sol 6.1, and this
62. 2:23one belongs to Opus 5.5. They are
63. 2:27identical. And as you can see here, we
64. 2:29have fresh new chat.
65. 2:32Um this is Sol, as you can see, it's on
66. 2:34lab, and Opus as well.
67. 2:37So, everything is good and ready. We're
68. 2:39going to use 6.1 Sol on the high
69. 2:42settings. All right, we're not going to
70. 2:44use Astro because we want to use
71. 2:46something affordable. Same thing on
72. 2:48Opus, we're going to use the 5.5. We're
73. 2:50not going to use 5.1.
74. 2:52Uh Fable 5.1, it is um it requires
75. 2:55usage, and we're going to set it on the
76. 2:57high settings. And as you can see, the
77. 2:59usage right here, the weekly limit is at
78. 3:0194%. We're going to see how much it's
79. 3:03going to take uh on both models. Opus is
80. 3:07set at 19%, and we're going to see how
81. 3:09much each one is going to take. And this
82. 3:11is going to happen in three rounds. So,
83. 3:14round one, they're going to build, and
84. 3:15they have to work at least an hour and a
85. 3:17half, so no rushing. Round two, we're
86. 3:20going to open the vault, which is the
87. 3:22out of sample, the data that they have
88. 3:24never seen. And we're going to see if
89. 3:26their candidates actually work there.
90. 3:28And finally, round three, whatever
91. 3:30survives, it's goes to a real broker.
92. 3:33We're going to go to MT5 tester on IC
93. 3:36Market. So, we see if it works or not.
94. 3:38So, let's get into it. So, here's the
95. 3:40first prompt, the exact same text into
96. 3:43both labs at the same moment. [snorts]
97. 3:45Here, we're going to set a goal. A goal
98. 3:47is just basically it's an objective that
99. 3:50it's going to keep pursuing. Um they
100. 3:52both have this feature. Same thing on
101. 3:55Office 5.5 and we're going to pass both
102. 3:58the prompt to both of them and we're
103. 4:01going to let them work.
104. 4:03Now, while they work, let's talk about
105. 4:05what's inside the lab.
106. 4:08Because this is not just a plain ChatGPT
107. 4:10chat in a browser tab, all right? If you
108. 4:12ask a normal AI for a trading strategy
109. 4:15and it's going to take 30 seconds and
110. 4:17you will get an RSI-based strategy with
111. 4:20a stop and a target. If you take it to
112. 4:22the lab and bang, that's it. It It will
113. 4:24fail miserably and I know that you have
114. 4:26tried it before. Well, Strategy
115. 4:28Architect works on works the other way
116. 4:31around. So, let me show you what a claim
117. 4:33you have probably heard.
118. 4:35When the Nasdaq gaps up at the open, the
119. 4:38gap gets filled. So, a normal AI, it
120. 4:41will just code it. It will trade the
121. 4:43gap, target the fill and done. The
122. 4:46Strategy Architect pulls it apart first.
123. 4:48So, what kind of gap? Is it big or is it
124. 4:51small? Is it after news or on a quiet
125. 4:54day? Who's trading at that moment and
126. 4:57why would they push price back? Also,
127. 5:00filled when? In the first hour or by the
128. 5:03close? Is it even better than a random
129. 5:05day? So, it compares a baseline and the
130. 5:07question most people never ask, what
131. 5:09happens on days the gap doesn't fill?
132. 5:12Because sometimes the failure can lead
133. 5:14to another strategy. So, it writes down
134. 5:17a dozen competing explanations like
135. 5:20these and makes its before it is allowed
136. 5:23to see a single result. And every test
137. 5:25is logged before the result comes in, so
138. 5:27it can't hide its failures. It can't
139. 5:29peek at the future.
140. 5:31And the vault, which is the
141. 5:32out-of-sample, stays locked until I open
142. 5:35it. And every mistake goes into the
143. 5:37memory its memory for the next campaign,
144. 5:39so we're not repeating the same mistake.
145. 5:41And this is just the tip of the iceberg.
146. 5:43There is a whole infrastructure
147. 5:45and a lot of moving parts below it.
148. 5:47We're going to talk about this in a
149. 5:48later video.
150. 5:50All right, so back in the labs and looks
151. 5:51like Opus is starting to finalize. So,
152. 5:55this is the right time for the
153. 5:56challenge. I'm going to send it
154. 5:59to Opus first because it looks like Saul
155. 6:01is still a bit behind. So,
156. 6:03I'm going to give it a little more time.
157. 6:05Now, why did I send that challenge
158. 6:08prompt?
159. 6:09Basically, what I learned is that AIs
160. 6:11sometimes get tunnel vision. So, they
161. 6:13grab the first decent idea and they're
162. 6:15going to keep polishing it and only
163. 6:17focusing on that. A challenge is going
164. 6:19to force them to stop, audit everything,
165. 6:22and zoom out.
166. 6:23Some of my best results came right after
167. 6:26this push.
168. 6:27Uh in my own work, I pushed more than
169. 6:29once, but
170. 6:31here is just one.
171. 6:33But still, uh it's the tip that worth
172. 6:35stealing. All right. Now, as you can
173. 6:37see, Opus finished its work in an hour
174. 6:40and uh and 20 minutes and basically
175. 6:43waited for the hour and a half mark to
176. 6:45seal
177. 6:46everything with it. It found three
178. 6:48candidates and Saul actually kept going.
179. 6:51So, we're going to wait for it to
180. 6:52finish. All right, so Saul finished
181. 6:54after about an hour and 40 minutes with
182. 6:57only one candidate.
183. 6:59And of course, all of them made money on
184. 7:01the data. So, Saul it found only one
185. 7:03strategy that made money. Uh Opus find
186. 7:06found three, but that's expected, to be
187. 7:08honest, and it means absolutely nothing.
188. 7:10All right, you can tune any strategy to
189. 7:12the past until it looks perfect. The
190. 7:15only question is whether it works on
191. 7:17data it has never seen. So, here's
192. 7:20prompt number two, which is permission
193. 7:22to open the vault, and every candidate
194. 7:24will be run on 2024 to 2026. Once, no
195. 7:28tweaking, no optimizing, no second
196. 7:31chances. All right, so the results are
197. 7:33in.
198. 7:33>> [clears throat]
199. 7:34>> Let's see.
200. 7:35Uh we're going to start with Saul and
201. 7:37its one and only candidate that won in
202. 7:39the in sample.
203. 7:41And it looks like that. It failed profit
204. 7:44factor 1.006,
205. 7:46basically break even. It had a 30%
206. 7:50drawdown from its highest equity curve.
207. 7:52So, it didn't just went into drawdown,
208. 7:55but it actually made money
209. 7:56and then started
210. 7:58losing it
211. 7:59along the way. But, look at the equity
212. 8:02curve. And this is interesting, to be
213. 8:04honest. So, through 2024 and 2025 on
214. 8:08unseen data, unseen years, it was
215. 8:11actually making money.
216. 8:14So, pretty consistent on unseen data.
217. 8:16And then 2026 comes,
218. 8:19the market changes, and everything
219. 8:22went to the drain. I don't know what
220. 8:24happened here. It needs more
221. 8:26investigation later on, but
222. 8:28this is strange.
223. 8:30Something changed in that particular
224. 8:33candidate or the mechanism that it
225. 8:35exploits. So, failed. Now, all this,
226. 8:38three candidates, um went to the out of
227. 8:41sample.
228. 8:42C1 passes, C3 passes, and C2 failed. So,
229. 8:47two candidates survived in the out of
230. 8:49sample.
231. 8:50And here's where it gets interesting. In
232. 8:53the final prompt, I told them to combine
233. 8:55every survivor into one book. So, for
234. 8:57example, C1 and C3 made money.
235. 9:00Uh and they trade very differently. So,
236. 9:02they have low correlation. So, one
237. 9:05has a bad month, the other usually
238. 9:07doesn't. So, when together, when you
239. 9:10merge them together, they can give us a
240. 9:12better equity curve and better metrics
241. 9:14than trading them separately,
242. 9:16you know, or choosing only one. So, as
243. 9:18you can see, [clears throat] on data
244. 9:19they had never seen, they made plus 65%
245. 9:22with a max drawdown of about 14%. More
246. 9:25return and a smoother ride, as you can
247. 9:27see. Um don't let the full history chart
248. 9:30fool you because I know it looks huge
249. 9:32between the book and the two strategies
250. 9:35separately, but mainly it is because of
251. 9:38the compounding effect.
252. 9:40Uh the unseen part right here is the one
253. 9:42that counts. So, Opus takes round two,
254. 9:45but I'm not calling it yet because
255. 9:47everything so far running Python and a
256. 9:50different uh data feed, real money
257. 9:53doesn't live in Python. So, we need to
258. 9:55test it on a a feed that we we can we
259. 9:57can actually trade. So, we translated
260. 10:00both strategies into MT5 Expert Advisor
261. 10:03and run them in IC Market on tester. So,
262. 10:06C3 is actually sensitive and the edge
263. 10:10lived in tiny details. So, price poking
264. 10:13just a hair above yesterday's high and
265. 10:16then coming back. Uh in Forex, every
266. 10:18broker has its own price feed. So, it's
267. 10:20not like futures where everyone gets the
268. 10:22same price from the exchange. And on C3
269. 10:25days, the two feeds only agreed 61% of
270. 10:28the time. So, the feed that I use in the
271. 10:31the internal backtesting, they're not
272. 10:33like huge differences, but they are only
273. 10:35small pips differences, but that's why
274. 10:38uh C3 failed because it used those
275. 10:41little pokes, tiny pokes as part of its
276. 10:43edge, you know? But, C1 was actually
277. 10:46bulletproof. So, as you can see, the
278. 10:47same days on both feed and is in IC
279. 10:50Market, it made plus 43
280. 10:53435%
281. 10:54from 2016 to 2026.
282. 10:57Uh the drawdown is bigger, there around
283. 10:5829% because you pay real commission. The
284. 11:02commission was a bit different from what
285. 11:03I modeled, but um
286. 11:05and and its curve has has these big
287. 11:07jumps. As you can see, you might think
288. 11:09that they're lucky days, but they're
289. 11:10not. That's uh the design and the logic
290. 11:13behind the strat behind the strategy.
291. 11:15So, every night, C1 will draw um the
292. 11:19Asian [clears throat] box, but it only
293. 11:21trades when the box is tight, tighter
294. 11:23than most of the last 20 nights. A tight
295. 11:26box means a tiny stop, usually 15 to 20
296. 11:30pips. And if the break is fake, it loses
297. 11:33that tiny stop but there but there's no
298. 11:36take profit, okay?
299. 11:38The strategy has no take profit. When
300. 11:39the break is real, it rides until noon,
301. 11:43New York, and gets out right before big
302. 11:46US news, all right? Uh it risks 12 pips
303. 11:49and sometimes it catches 93 pips and in
304. 11:52a lot of instances. That's eight times
305. 11:55the risk or eight R or seven R trades.
306. 11:58So, days like this can pay for eight
307. 12:00losers and it's not a fluke or lucky
308. 12:03days. C1 had four to nine of these big
309. 12:06days every single year from 2016 to 2023
310. 12:11and it kept coming in 2024 and 2025 and
311. 12:142026 on seen data. So, it is a pattern
312. 12:18in price, it is not luck, it is
313. 12:20consistent.
314. 12:22So, the final verdict, office wins all
315. 12:24three rounds. It's strategy, the first
316. 12:26candidate survived data it had never
317. 12:28seen and it survived a real broker, so
318. 12:31broke even with a 30% drawdown
319. 12:33and never made it to round three and um
320. 12:37now my honest opinion, would I trade
321. 12:39this one strategy on one pair alone? No.
322. 12:42And nobody should. One strategy is just
323. 12:44one brick. But put it in a book of 20
324. 12:48uncorrelated strategies, some on
325. 12:49USD/JPY, some on US 100, some on EUR/USD
326. 12:53from different families like mean
327. 12:55reversion, momentum, breakout,
328. 12:57pre-breakout, whatever. And when one
329. 13:01struggles, the others will carry it.
330. 13:04So, that's how you get a steady equity
331. 13:07curve, a smooth one.
332. 13:09And at the end, if you want the PDF of
333. 13:11the winning strategy with the full rules
334. 13:12and the logic, guys, just remember, all
335. 13:15you have to do is to subscribe and leave
336. 13:17a comment down below and I'll send it to
337. 13:19you.
338. 13:19Now, happy trading and good luck.