Opus 4.8 + Claude Code + MCP = King of Algo Trading! — Transcript

Full transcript

1. 0:00In this video, I'm going to show you how
2. 0:01to use Claude Opus for 4.8, which has
3. 0:04just been released for algo trading in
4. 0:06Python. This model is supposed to be the
5. 0:09best one that the money can buy at the
6. 0:10moment and is supposedly the king. And
7. 0:13we're going to use it to write trading
8. 0:14strategies, to backtest them, to iterate
9. 0:16over them until it improves the results.
10. 0:19And not only that, we will also ask the
11. 0:21agent to ensure that the entry rules of
12. 0:23the strategy that it has wrote for us
13. 0:25has some statistical meaning and that
14. 0:28they are not pure noise or some lucky
15. 0:30results. And in the end, I will also go
16. 0:32through my process of running the Monte
17. 0:34Carlo simulation in order to decide
18. 0:37whether or not the results that I'm
19. 0:38getting is good enough for running it
20. 0:41live. And before we continue, I got to
21. 0:43say I'm not a financial advisor and this
22. 0:45video is for educational only. All
23. 0:47right, I'm super psyched, so let's get
24. 0:49right into it.
25. 0:52Now, one question that you need to
26. 0:53answer is how you're going to get access
27. 0:55to Opus 4.8, because this is one of the
28. 0:57most expensive models at the moment and
29. 1:00it is not affordable just for everybody.
30. 1:02However, my favorite way of using it is
31. 1:04through the Claude Code subscription,
32. 1:06which give us generous limits. However,
33. 1:08we do need to talk about like which plan
34. 1:10you need to get. And realistically, if
35. 1:12you want to use Claude Opus and not hit
36. 1:14the limits like right away, you're going
37. 1:16to need the max plan. Now, there are two
38. 1:18max plans, $100 and $200. I am on the
39. 1:21$100 plan and I couldn't be happier
40. 1:24about it. I have never hit the limits
41. 1:26and it is perfect. But if this isn't
42. 1:28affordable for you, yes, you can go with
43. 1:30the pro plan, but chances are that you
44. 1:32will hit the limits. Now, there are
45. 1:33multiple ways to use Claude Code. One is
46. 1:35to use their desktop app and the other
47. 1:37is to use it via the terminal. And I'm
48. 1:39using the terminal one because I find it
49. 1:41to be a little bit faster. So, first
50. 1:43things first, we're going to use the
51. 1:44Jesse framework for the algo trading
52. 1:46part, which will also expose an MCP,
53. 1:48which we can use to connect to Claude
54. 1:50Code. And the way I'm going to run
55. 1:51Claude Code is via the terminal. Before
56. 1:53I do that, we're going to have to use
57. 1:55the MCP which Jesse exposes. So, if I
58. 1:57open my terminal and simply run Jesse
59. 2:00run, it's going to start the Jesse
60. 2:01dashboard. And here's the address for
61. 2:03the dashboard itself and here's the
62. 2:04address for the MCP. So, this is going
63. 2:06to be the one that I'm going to use.
64. 2:07Now, if you check out the documentation
65. 2:08for Jesse's MCP under the connect in
66. 2:11cloud code, you can see this is the sole
67. 2:14command that we're going to have to run.
68. 2:16So, inside my terminal, I'm going to
69. 2:17have to paste this command. And if you
70. 2:18remember, for me, this port was a little
71. 2:20bit different because I had changed it
72. 2:22inside the .env file of my project. So,
73. 2:25it's 9,000. So, it's going to be like
74. 2:26this. I've already done this, so I don't
75. 2:28have to do it again. And that's it,
76. 2:29guys. Now, I am ready to run the cloud
77. 2:31code inside the terminal, which is
78. 2:33inside the directory which I have Jesse
79. 2:35installed. Next, I'm going to ensure
80. 2:37that I have selected the Opus 4.0 as the
81. 2:40model for this session. Now, before we
82. 2:42continue, I want to ensure that the
83. 2:44agent is going to follow the agent.md
84. 2:47file of my project, which is basically
85. 2:49telling the model how to behave. Now,
86. 2:51this file is shipped with every Jesse
87. 2:53project, and whenever you update your
88. 2:55Jesse, it will also update this file
89. 2:56automatically. So, you do not need to
90. 2:58worry about this much. We can just close
91. 3:00it. And here, I'm going to simply attach
92. 3:02it to my current session. Next, I'm
93. 3:04going to simply say, "Write a trend
94. 3:05following a strategy for 2024 on the
95. 3:08BTC/USDT
96. 3:10market. Once you write the strategy, run
97. 3:13execute a rule validation test to ensure
98. 3:16the entry rule of the strategy has
99. 3:18statistical significance and isn't
100. 3:20random noise. If it failed, tell me
101. 3:22about it and then continue iterating
102. 3:24over the entry rules until you get it
103. 3:26right. Once it does pass the test, run a
104. 3:28backtest on it and show me the results."
105. 3:29So, here's the thing, guys. I'm not
106. 3:31asking the agent to do everything all at
107. 3:33once. I could simply tell it to continue
108. 3:36iterating over until the backtest also
109. 3:38looks great on that period. But, I'm not
110. 3:40going to do that because I want to show
111. 3:41you how you can utilize this agent thing
112. 3:44to develop your own strategies
113. 3:45step-by-step. So, even if you, for
114. 3:47example, don't really know what a
115. 3:48backtest is, or if you're not getting
116. 3:50good results in the backtest and you're
117. 3:52wondering what to do next, we can still
118. 3:54use this thing. All right, let's hit
119. 3:55enter and see what it does. All right,
120. 3:56so it decided to create a Donchian
121. 3:59breakout strategy plus the EMA for the
122. 4:02trend and it's saying that the long
123. 4:03entry is going to be whenever the candle
124. 4:05closes above the prior breakout period
125. 4:08high and the close is bigger than the
126. 4:10current trend period using the EMA. It's
127. 4:13creating the significance test. It's run
128. 4:14it and it says that it's passed. So, the
129. 4:16P value is almost zero, so it is
130. 4:18definitely less than this value which is
131. 4:20our threshold, which means that the
132. 4:22entry rule does indeed have statistical
133. 4:24significance. And here's what I meant
134. 4:26earlier. Even if you do not know how to
135. 4:28read the results of a significance test,
136. 4:31you can read the guidelines that the
137. 4:32agent wrote for you. So, for instance,
138. 4:34it's telling me that the Donchian
139. 4:36breakout plus EMA trend entry is not
140. 4:38random noise. It beats every one of the
141. 4:412,000 random entry variants. So, because
142. 4:44it did this, we can conclude that the
143. 4:46entry rule of the strategy was not
144. 4:48random noise. Now, if you do not
145. 4:50understand how this test works, that's
146. 4:51okay. I'm going to create a video about
147. 4:53it in the future, but for now, what's
148. 4:55important is you learn how to use it and
149. 4:58the fact that even if you do not know
150. 5:00that, the agent is able to do it for you
151. 5:02is absolutely amazing. Next, you can see
152. 5:04that the model did the position sizing
153. 5:06of the strategy because you see so far
154. 5:08it only had to written the entry rule of
155. 5:10the strategy. It's only after this,
156. 5:12after ensuring the entry rule is not
157. 5:15pure noise and it does have some edge,
158. 5:17that's when it decided to continue to
159. 5:19write the position sizing and the exit
160. 5:21rules of the strategy. Next, it is
161. 5:23creating the back test for BTC on the
162. 5:2630-minute time frame for the full year
163. 5:28of 2024. It is running the back test. It
164. 5:30keeps checking the results and then it
165. 5:32says the results already and here we
166. 5:35should be able to read the results. Now,
167. 5:36by the way, it is also writing a report
168. 5:39inside this file. So, even if you want
169. 5:41to see what exactly the model did later
170. 5:43and you lost access to these logs, you
171. 5:45can absolutely do that. All right, so
172. 5:47here is the results for that rule
173. 5:49significance test again, and here's the
174. 5:51full backtest results. So, we can see
175. 5:53the net profit is absolutely negative.
176. 5:55The total trades is a huge number, 208,
177. 5:58and the average win to loss ratio is
178. 6:002.20, which is typical in a
179. 6:03trend-following strategy. All right, so
180. 6:04the results are negative so far. So,
181. 6:06let's see if we can ask the model to
182. 6:08improve it. It's also giving me its
183. 6:10honest takeaway, which is really nice.
184. 6:12And this is actually an an important
185. 6:13thing, which is strategies in trade has
186. 6:16an edge, it doesn't mean that the
187. 6:17strategy is going to print money. In
188. 6:19this case, for example, the strategy is
189. 6:21actually losing money because the
190. 6:22strategy is executing too many trades.
191. 6:25Now, this is a classic thing in
192. 6:26strategies, and whenever this happens,
193. 6:28usually I add a couple of filters in
194. 6:30order to improve the results and, you
195. 6:32know, take the less amount of trades and
196. 6:34pay less amount of trading fees. Now, in
197. 6:36the end, it's actually asking me if I
198. 6:37want to iterate the full strategy in
199. 6:40order to improve the results. And that's
200. 6:41exactly what I want to do, so I'm going
201. 6:43to ask the model to do that. Please
202. 6:44iterate over the results three more
203. 6:46times and try to improve it, and then
204. 6:48compare the results and show them to me
205. 6:50to see which one it looks better. And
206. 6:51whenever you do, first run a
207. 6:53significance test to ensure that new
208. 6:55iteration not pure noise, and it does
209. 6:57have an edge before continuing to the
210. 6:59backtest. All right, so while this is
211. 7:00going, let's take a look at here. So,
212. 7:02the strategies folder, we can see a new
213. 7:05strategy I've been created called BTC
214. 7:07Trend 30 minutes, and inside it we have
215. 7:08this init file, which is where I define
216. 7:11the properties of the strategy. You can
217. 7:13see how it defined the Donchian Channel,
218. 7:15trend EMA, the ATR. Here are the entry
219. 7:18rules of the strategy. So, for example,
220. 7:19for the short long, it's simply saying,
221. 7:21"I'm going to go long whenever the
222. 7:23closing price is above the prior high,"
223. 7:25which it is using the Donchian Channel
224. 7:27again, "and the current closing price is
225. 7:29above the current trend EMA value." For
226. 7:32a short entry, it's doing the opposite.
227. 7:34And here we can see how it's doing the
228. 7:36position sizing. So, it's simply saying
229. 7:38that for the entry, I want to use the
230. 7:39current price, which is a market order.
231. 7:41For the stop, it's using the current
232. 7:43entry minus the current ATR multiplied
233. 7:46by a stop ATR multiplier. So, let's see
234. 7:48what this value is. Okay, so by default
235. 7:50it is set to two. And next for the
236. 7:52quantity, it is using the risk to
237. 7:53quantity utility function of Jesse. It's
238. 7:55passing the current available margin,
239. 7:57the risk percentage, which it has set to
240. 7:592%, and the entry, the stop price, which
241. 8:02we calculated here, and then it is
242. 8:04setting the fee price. So, this is
243. 8:05absolutely correct. Then it is
244. 8:06submitting the buy order using the
245. 8:08quantity and the entry, and it is also
246. 8:10submitting the stop loss here. Now, it
247. 8:12did not need to define this since we are
248. 8:13using a market order, so there's not
249. 8:15going to be any cancellation for the
250. 8:17entry. And then the update position
251. 8:19function, which is what we execute
252. 8:20whenever a new candle closes while we
253. 8:23have an open position. So, it's saying
254. 8:25that if it's a long position, the exit
255. 8:27price is going to be the Donchian
256. 8:29Channel's lower band, and if the closing
257. 8:31price is below that, then we want to
258. 8:33liquidate the current position. And for
259. 8:34a short position, obviously it's doing
260. 8:36the opposite. That's it. So, this is
261. 8:37pretty simple a strategy, and I would
262. 8:39say a classic one. All right, let's take
263. 8:41a look at what the agent did so far. So,
264. 8:44it looks like it defined three variants,
265. 8:46so it's calling them V2, V3, and V4.
266. 8:48Next, it is also comparing the results
267. 8:50of their RST test, which again is a rule
268. 8:53significance test, and all of them are
269. 8:55passing. Now, by the way, if you take a
270. 8:57look at this value here, you can see the
271. 8:59analyzed return of them are different,
272. 9:01but I wouldn't necessarily say that like
273. 9:04one is better than the other because of
274. 9:06this. Okay, so all that I need to see
275. 9:08with the RST test is whether or not the
276. 9:10strategy's entry has some edge, so I can
277. 9:12continue to define the exit rules of the
278. 9:15strategy. That's it. Because if it tells
279. 9:17me that it doesn't, it could just mean
280. 9:19that my entry rules are just some lucky
281. 9:21guess. So, for example, if I have a
282. 9:23strategy that always goes long, of
283. 9:25course during a bull run it's going to
284. 9:26make me money, but it doesn't mean that
285. 9:29it does actually have some predictive
286. 9:31power. All right, it finished the
287. 9:32results, and here's again the table for
288. 9:34the RST test, and here's the table for
289. 9:37the backtest results. And we can see
290. 9:39this comparing the V1 of the strategy,
291. 9:41V2, V3, and V4. So, four different
292. 9:43variants. It's telling me which one
293. 9:45looks better. So, the variant three is
294. 9:47actually profitable. It has a sharp
295. 9:49ratio of 2.11, which is absolutely
296. 9:51amazing. The max drawdown is only minus
297. 9:5417%, which is great. The number of
298. 9:56trades is 86, which actually makes a lot
299. 9:58of sense. The win rate is low, which
300. 9:59makes sense for a trend following
301. 10:00strategy, especially if the average win
302. 10:03to loss ratio number is high. Which we
303. 10:05can see here, it is near four. So, of
304. 10:07course, this makes sense. And here's
305. 10:09another cool thing that was added to
306. 10:10Jesse's MCP recently. After it is done,
307. 10:13it will also print out the URL to the
308. 10:15dashboard. So, for example, for the V3,
309. 10:17I can click on this URL, and it will
310. 10:19open the dashboard of Jesse, so I can
311. 10:22look at some charts. So, for example,
312. 10:24here's the equity curve. Here's a log
313. 10:25scale version of it, and here is the
314. 10:27five worst drawdown periods of the same
315. 10:30strat. And here is the monthly returns
316. 10:31of the strategy. So, as you can see,
317. 10:33even though it is doing absolutely
318. 10:35awesome in three months, in some other
319. 10:37months, it's actually negative, which is
320. 10:39a typical thing with a trend following
321. 10:40strategy. You can see the agent is also
322. 10:42writing some summaries of like what the
323. 10:45strategy's intro rules are about, what
324. 10:47it's doing, and things like that. And if
325. 10:49I want to, I can also click on this and
326. 10:51view the interactive charts. So, for
327. 10:53example, I can take a look at the trades
328. 10:56that we took. So, for example, in this
329. 10:57one, if I click here, it will take me to
330. 10:59when the position was opened and when it
331. 11:01was closed. So, as you can see, yes,
332. 11:03this was a clear uptrend moment. And it
333. 11:05makes sense that in these one, it's just
334. 11:07losing money because this is a ranging
335. 11:08period. Let's take a look at another
336. 11:10one. So, this one, for example, it went
337. 11:12long here, and it closed it here. Let's
338. 11:15take a look at the losing one, such as
339. 11:16this one. So, you see, it opened it
340. 11:19here, but it closed it here. And by the
341. 11:21way, because it opened at the closing
342. 11:23price of this candle and at the closing
343. 11:25price of this one, so that's why we
344. 11:27actually lost money. Because if we had
345. 11:29entered it here and closed it here, it
346. 11:31would have been a tie, or maybe a small
347. 11:33win. All right, so back to the terminal.
348. 11:35So, it's telling me that the V3 actually
349. 11:37wins. It has the highest net profit and
350. 11:39the best risk-adjusted metrics such as
351. 11:41the sharp ratio. So, it basically
352. 11:43concluded everything that I just said.
353. 11:44So, even if you do not know how to read
354. 11:47a sharp ratio, the agent does. Now, by
355. 11:49the way, the strategy that it wrote
356. 11:51actually looks really good. So, I'm
357. 11:53happy with the results of the Opus 4.8
358. 11:55model so far. And it's also telling me
359. 11:57what fixed it results. So, what actually
360. 11:59fixed it is that the entry was never the
361. 12:02problem. So, the RC proved it every
362. 12:04time. We won blood out from shorts in a
363. 12:07bull year. A two-tight exit plus fee
364. 12:09drag. Removing shorts and widening the
365. 12:11exit flipped from this value into a
366. 12:14positive one. A slower candle breakout
367. 12:16added the rest. However, as you probably
368. 12:18know, we cannot trust the results of the
369. 12:20backtest just yet and go to live trading
370. 12:23with it. We also need to run a Monte
371. 12:25Carlo test at the very least, of course,
372. 12:27to make sure the results of the strategy
373. 12:29is not overfit. And in fact, that's what
374. 12:31the agent is also telling me. So, one
375. 12:33honest caveat. 2024 was a strong
376. 12:35uptrend, which flatters a long-only
377. 12:37trend follower. Want me to Monte Carlo
378. 12:39V3 overfit check and validated out of
379. 12:42sample on 2023 and 2025 before you start
380. 12:45these numbers. Now, we could ask the
381. 12:47model to also run the Monte Carlo, and
382. 12:49that's exactly what I did, but there was
383. 12:51a problem with it. So, I'm going to try
384. 12:53running it manually on the dashboard of
385. 12:55Jesse. So, here's the exchange that I've
386. 12:56chosen, the BTC USDT route, and let's
387. 13:00change this into 30 minutes. I have
388. 13:02chosen the BTC trend 30 minute V3
389. 13:04strategy, which is what the agent made
390. 13:06for us, and I have picked the entire
391. 13:082024. Now, we are ready to start the
392. 13:11Monte Carlo test. So, the Monte Carlo
393. 13:13trades is over. So, this one is always
394. 13:16really fast, and what it basically gives
395. 13:18us is what would have happened if the
396. 13:20order at which we took the trades were
397. 13:22different. So, for example, if instead
398. 13:24of beginning with the winning trades, if
399. 13:26we had with losing ones. Like, if that
400. 13:28was the case, what would have been the
401. 13:30max drawdown of the strategy? And in
402. 13:32that case, were we prepared for it? So,
403. 13:34basically, this is for the position
404. 13:36sizing of the strategy. And it's telling
405. 13:38me that in the worst 5%, my max drawdown
406. 13:41would have been minus 23%. But, in the
407. 13:43original backtest, it was only minus
408. 13:4517%. We could have also been luckier,
409. 13:47like in the best 5%, and in that case,
410. 13:50it would have been minus 8.5%.
411. 13:52But, the thing that I really want to
412. 13:53know is the Monte Carlo candles, which
413. 13:55is where we execute a Monte Carlo
414. 13:57simulation by using synthetic data,
415. 13:59which has been created based on the
416. 14:02original data of the market. So, this
417. 14:04one is basically like a stress test for
418. 14:07the strategy in order to see how robust
419. 14:10it is, which also tells us whether or
420. 14:11not it is overfit or not. So, as you can
421. 14:14see, this one's taking longer, which
422. 14:15makes perfect sense. Now, while that's
423. 14:17going, I want to quickly remind you guys
424. 14:19about our Telegram. It's the fastest way
425. 14:21to get notified about my future work,
426. 14:23whether it's a new tutorial or that I
427. 14:25create. Also, don't forget to check out
428. 14:27our free Discord, where more than 5,000
429. 14:29members like you and I are hanging out
430. 14:31there and helping out each other with
431. 14:32algo trading, so we can all succeed
432. 14:34together. The links for both are down in
433. 14:36the description. All right, the candles
434. 14:37Monte Carlo is over. Here, we can see
435. 14:40the chart. The yellow line is the
436. 14:42original backtest, and these blue lines
437. 14:44are the simulations. Now, the fact that
438. 14:46the original backtest lines is almost in
439. 14:48the middle is a good sign. If we take a
440. 14:50look at the results table, we can see
441. 14:52that the original backtest sharp ratio
442. 14:55was 2.48, while the best 5% was 3.41.
443. 14:58And the fact that this one's lower than
444. 15:00this is a good sign. So, I wouldn't say
445. 15:02the results are overfit. However, the
446. 15:05median number is 1.87, and this one is
447. 15:09lower than the one for the original
448. 15:11backtest. So, the original backtest
449. 15:13results is somewhere between these two
450. 15:15numbers. So, again, I wouldn't say it's
451. 15:17overfit, but I wouldn't say it is like
452. 15:19super robust, either. So, that's it,
453. 15:21guys. This is how I personally read the
454. 15:23Monte Carlo results in order to take a
455. 15:26guess at whether or not the results of
456. 15:29the strategy are going to be as good in
457. 15:31the live trading as what I'm seeing in
458. 15:33the back test. However, let's go to the
459. 15:36history because I want to show you the
460. 15:38results of the week three in 2024. And
461. 15:42also, I'm going to add two more. So,
462. 15:45this one being in 2023. Let's ensure the
463. 15:49fast mode is on. I don't need the
464. 15:51interactive chart anymore.
465. 15:53All right, so let's run this. And I'm
466. 15:55going to run another one this time for
467. 15:572025. So, let's run it. There we go. So,
468. 16:00here's the result for 2023. And as you
469. 16:03can see, again it is profitable. So,
470. 16:05it's making 62%. All right, so let's go
471. 16:08to 2025. The strategy is falling apart
472. 16:10if we are not in a clear uptrend. This
473. 16:13by itself could be enough reason to not
474. 16:16jump into trading this strategy live
475. 16:18yet. I would personally say that it is
476. 16:20not ready for production and it needs
477. 16:23further work because even if you are not
478. 16:25in a clear uptrend, I would prefer the
479. 16:27strategy not to at least lose a lot of
480. 16:29money in that period. Or of course, we
481. 16:30could also develop a short only strategy
482. 16:33to run besides this one to make up for
483. 16:35the losses of this one. Now, I am going
484. 16:37to submit the result of the strategy to
485. 16:40our strategies page where you can see
486. 16:42and browse other strategies that I've
487. 16:44developed before. If you click on them,
488. 16:46you can see the results of their metrics
489. 16:48on different periods and symbols or time
490. 16:50frames. So, feel free to check out that
491. 16:52page as well. And that's it, guys. I
492. 16:54hope you enjoyed this video and the
493. 16:55experiment that we just did. And don't
494. 16:57forget that the whole point of this was
495. 16:59not to give you one strategy. It was to
496. 17:01give you a tool which can save you
497. 17:03countless hours when developing and
498. 17:06evaluating your own strategies. All
499. 17:08right, let's pick the winner from the
500. 17:09previous video. The winner is as always
501. 17:11solid. Thank you so much for your
502. 17:12comment. Please reach out to me so I can
503. 17:14send you your bank tokens. Thank you so
504. 17:16much for watching. I'll see you in the
505. 17:17next one.
506. 17:22>> [music]