Sonnet 5 + Claude Code strategy makes 369% — Transcript

Full transcript

1. 0:00The Anthropic team just released Claude
2. 0:02Sonnet 5. Claude Sonnet used to be my
3. 0:04favorite model of all time because it
4. 0:07gave me the best quality comparing the
5. 0:09results. Especially in the past where I
6. 0:11would just wanted the $20 subscription
7. 0:13model of Anthropic, Claude Sonnet was a
8. 0:15really good deal for me. So, I'm really
9. 0:17excited to try out this new version. As
10. 0:19you guys know, on this channel what I
11. 0:21care for is the performance of the model
12. 0:24specifically for trading. I want to see
13. 0:26how good of a strategy it can write for
14. 0:28me and how good it is at following
15. 0:30instructions to do real research for me.
16. 0:33That includes running the backtest,
17. 0:34doing optimization to find the best
18. 0:36parameters of the strategy, and running
19. 0:38Monte Carlo sessions to ensure the
20. 0:40results that it finds for me aren't
21. 0:41overfit. Now, before we continue, I got
22. 0:43to say I am not a financial advisor and
23. 0:45this video is for educational purposes
24. 0:47only. So, with that out of the way,
25. 0:49let's get right into it.
26. 0:54So, here's a blog post by the Anthropic
27. 0:56team and as you can see, here's the
28. 0:59benchmark So, for agent coding, which is
29. 1:01exactly what we need, it is a scoring
30. 1:0363%
31. 1:05which is almost 5% above Sonnet 4.6,
32. 1:09which was a really good model by the
33. 1:11way. Like, I've been using it until
34. 1:12yesterday. And Opus 4.8, which is the
35. 1:16easiest and my main driver because I
36. 1:18have a max subscription plan of
37. 1:20Anthropic, is still beating it. So, it
38. 1:23is 69%. So, let's not forget this. And
39. 1:26the same really goes for other stuff.
40. 1:27So, in this one, it is just 2% below
41. 1:30Opus 4.8, but it is beating Claude 4.6
42. 1:34by significant margin, actually. And in
43. 1:37the knowledge work, which is also
44. 1:39important because we can consider this
45. 1:42one for how good the model is at, for
46. 1:44example, with its trading knowledge. So,
47. 1:47not just following the tools or doing
48. 1:49agent work for us, but it's actual
49. 1:52knowledge. Like, what sort of indicators
50. 1:54should it use? Or if it's not doing
51. 1:56well, let's say in a range market, what
52. 1:58should it do? So, that sort of thing.
53. 2:00But, guys, I think the main reason that
54. 2:02Anthropic didn't release the model that
55. 2:04is actually beating Opus, because that's
56. 2:06how they used to do it in the past, is
57. 2:08because they were worried that the model
58. 2:10is going to get flagged just like Claude
59. 2:11Fable, so they didn't want it to get
60. 2:14banned. I'm guessing that's why they are
61. 2:16doing it this way, but overall, the fact
62. 2:18that it's doing better than Sonnet 4.6
63. 2:20is a good sign and enough for me. Now,
64. 2:23as always, guys, I'm using framework for
65. 2:26doing the algo trading side of things.
66. 2:27So, if you haven't installed it already,
67. 2:30now's a good time to do so. Here's the
68. 2:32documentation where you can find
69. 2:33everything that you need. And next, I'm
70. 2:35going to use the Claude code inside the
71. 2:37terminal, which is running inside my
72. 2:39code editor. And the first step is to
73. 2:41ensure the Jesse MCP has been added to
74. 2:44your Claude code. For that, I'm simply
75. 2:46going to copy this address from here and
76. 2:48paste it here. Now, you need to be
77. 2:49careful with this address. You see, it
78. 2:51says using the port 9002, but for my
79. 2:54project, actually running on port 9007.
80. 2:57But, it doesn't matter, because I have
81. 2:59already added this before, and Jesse MCP
82. 3:01is connected to my Claude code. And just
83. 3:03to ensure, I will simply say, "Hey, can
84. 3:06you access Jesse MCP?" And yes, it says
85. 3:10Jesse MCP is connected and working. Now,
86. 3:12by the way, guys, the Sonnet 5 is
87. 3:14available and it's been selected here.
88. 3:16And if you haven't already switched to
89. 3:18it, you're going to need to run {slash}
90. 3:20model and then pick Sonnet, which is now
91. 3:23using Sonnet 5. All right, so, to prompt
92. 3:25the agent. So, I'm simply going to say,
93. 3:27"Hey there, do research and find me two
94. 3:29trading strategies. Each one of them
95. 3:31should have a sharp ratio above one in
96. 3:33the last 4 years. They should be trend
97. 3:36following, include both long and short
98. 3:38positions, and risk 3% of the account's
99. 3:40capital per each trade. In the first
100. 3:42strategy, once the position is open,
101. 3:45close it using a trailing stop. In the
102. 3:47second strategy, close the position at a
103. 3:49specific points determined by the ATR
104. 3:52indicator. You can test them on BTC,
105. 3:54ETH, and SOL on the hourly timeframe.
106. 3:57Every time you develop a strategy,
107. 3:59validate the result using a statistical
108. 4:01significance test before writing the
109. 4:03full strategy. Only proceed if the
110. 4:05strategy's entry rules demonstrates
111. 4:07genuine statistical significance. Feel
112. 4:10free to use optimization to improve the
113. 4:12results. At the end, apply Monte Carlo
114. 4:14simulations to ensure the results are
115. 4:16not overfit. Continue until you find
116. 4:18strategies that fully meet the criteria
117. 4:21requested. Do not prompt me in the
118. 4:23meantime. Good luck. And that's it. Now,
119. 4:25because I am pasting basically a lot of
120. 4:28content all at the same time, I'm seeing
121. 4:29it like this. But if I hit enter, we can
122. 4:32see it all being applied here.
123. 4:34Everything looks good and we can just go
124. 4:36and let Claude do his thing. Now, one
125. 4:38thing I really liked about this model is
126. 4:40that it is significantly faster. Whether
127. 4:43it is actually worth it to use this
128. 4:44comparing to Opus, I'm not sure at this
129. 4:47point. We're going to see, I guess, but
130. 4:49I do know that it is a lot faster to
131. 4:50run. So, anyways, let's go and come back
132. 4:52later. All right, so it's been around 1
133. 4:55hour and the agent is done. So, if you
134. 4:57come lower, we can see that it has been
135. 5:00calling the MCP multiple times. It first
136. 5:03checked and ensured that the necessary
137. 5:04candles for ETH and SOL are present in
138. 5:07my database and apparently a little bit
139. 5:10of BTC was missing, which it re-imported
140. 5:12for me. Then it wrote the strategies and
141. 5:14ran them simultaneously. And then it ran
142. 5:16an optimization session to improve the
143. 5:19results. And then it kept doing this and
144. 5:21it found that most of them are overfit
145. 5:23results. But it didn't give up and
146. 5:25continued doing research. Long story
147. 5:27short, at this point in time, it did
148. 5:29find one strategy which was way above
149. 5:32the criteria that I set for the agent.
150. 5:35So, it had a sharper 1.52.
151. 5:37But another one was almost as good as
152. 5:40what I wanted, but not quite. So, the
153. 5:42agent was being lazy here, which isn't
154. 5:44really great. And then the agent was
155. 5:46like, "Yeah, I'm done." Like I did his
156. 5:48job, but it hadn't. So, first of all, I
157. 5:51asked him to give me the URL for the
158. 5:53dashboard so I can check them out
159. 5:54myself, but I also specifically said
160. 5:56that one of them hasn't met the target
161. 5:58yet. Why did you stop? So, after I was
162. 6:01being hard on the agent, it did continue
163. 6:03working until this point in time, which
164. 6:06it did indeed reach the target that I
165. 6:08needed. And it also gave me the URLs for
166. 6:11the dashboard so I can check the results
167. 6:13both for the backtest and the Monte
168. 6:15Carlo simulation in order to ensure the
169. 6:17result isn't overfit. So, let's open
170. 6:19this one for the backtest first, and
171. 6:21then this one for the Monte Carlo. Now,
172. 6:24while that's going, I want to quickly
173. 6:25remind you guys about our Telegram. It's
174. 6:27the fastest way to get notified about my
175. 6:29future work, whether it's a new tutorial
176. 6:31or a tool that I create. Also, don't
177. 6:33forget to check out our free Discord
178. 6:34where more than 5,000 members like you
179. 6:36and I are hanging out there and helping
180. 6:38out each other with algo trading so we
181. 6:40can all succeed together. The links for
182. 6:42both are down in the description. So,
183. 6:44look at this. This is amazing, guys.
184. 6:46Like this is if we compare the results
185. 6:48to the benchmark, which is like other
186. 6:50assets that we were trading, but this
187. 6:52isn't really a good way to compare it
188. 6:54because it clearly the position sizing
189. 6:56of the strategy isn't enough. We could
190. 6:58increase it in order to actually give us
191. 7:00better returns than the benchmarks, but
192. 7:03if you consider the volatility, this one
193. 7:05is clearly winning. Now, let's scroll
194. 7:07down. So, in this chart, which is the
195. 7:09max drawdown, and there aren't any
196. 7:11benchmarks here, we can clearly see that
197. 7:13this is a smooth equity curve, and it
198. 7:16looks amazing. And here we can see the
199. 7:18monthly returns. You see, it's green on
200. 7:20most months, and it is definitely green
201. 7:23every single year. But the reason these
202. 7:25numbers aren't like huge is because the
203. 7:27max drawdown of the strategy is only
204. 7:29minus 7%. So, this is really low. So, it
205. 7:32made 75%. This can easily be increased.
206. 7:35It's executed 1,234
207. 7:39trades, which is amazing. The win rate
208. 7:41was 36%. so this is something to be
209. 7:44aware of because that means
210. 7:45approximately like 6 and 1/2 out of
211. 7:47every 10 trades are going to be losing
212. 7:49ones. So, this is perfectly normal for a
213. 7:51trend following strategy because the
214. 7:53wins that you have are going to be
215. 7:55significantly bigger than the losses. In
216. 7:57fact, if we take a look at this value
217. 7:59here, which is the average win to loss
218. 8:01ratio, it is 2.24. This is why even
219. 8:04though the win rate is way below 50%, we
220. 8:07are being profitable overall. The Sharpe
221. 8:08ratio is 1. 53. So, overall, this is
222. 8:11great and it is executing on three
223. 8:13symbols simultaneously. So, this is also
224. 8:15another reason why we're getting good
225. 8:17results. And in order to beat this, so
226. 8:19this is what I can do. I could just
227. 8:21click on new session and go and edit the
228. 8:24strategy myself and go to the position
229. 8:27sizing part, which is here. So, you see
230. 8:29this is the quantity of the order,
231. 8:31right? So, I could easily multiply this
232. 8:33by something like three, which by the
233. 8:36way, guys, I don't suggest you do this
234. 8:37for production because you want to be
235. 8:40completely aware of how much you are
236. 8:42risking. So, this isn't exactly how I
237. 8:44would do it in production, but just in
238. 8:46order to demonstrate this in this
239. 8:49backtest, it's a good solution. So,
240. 8:51let's save this, go back, and give it
241. 8:53another shot. There we go. So, now you
242. 8:55see in the equity curve of the
243. 8:58portfolio, which is this purple one, we
244. 9:00are beating all the other stuff even P&L
245. 9:03wise. So, this is pretty great to see,
246. 9:05but of course, when we do that, the max
247. 9:07drawdown also increases. So, it isn't
248. 9:10minus 7% anymore, it is minus 20%, which
249. 9:12is is still way acceptable for me. But
250. 9:15the yearly returns have significantly
251. 9:17increased now. We even have a year with
252. 9:1986%. We have another one with 67, so
253. 9:22this is pretty good. And if you want to
254. 9:24see some of the individual trades, we
255. 9:26can click on this view chart button
256. 9:28here. And here is the complete list of
257. 9:30the trades that it took. So, for
258. 9:31example, this one looks good, so let's
259. 9:33click on it. All right, so it went short
260. 9:35and then closed the same trade on the
261. 9:37same candle. In this losing one, again
262. 9:40the same thing happened. Here we went
263. 9:42long and then close it on the same
264. 9:44candle. Here, here we went short and
265. 9:48then we close it here. We can also
266. 9:50change this into ETH to see the trades
267. 9:53that we took for ETH. So, we went short
268. 9:55here and then close it here. We went
269. 9:57short here and close it here. And then
270. 10:01short here and close it here. So,
271. 10:02overall, I actually like the results
272. 10:04better for ETH. So, this looks pretty
273. 10:06good. And with this new position sizing
274. 10:08we are making 369%
275. 10:12but this is over 4 years. Now, another
276. 10:14thing is the Monte Carlo simulation that
277. 10:16we have to check out, which is how we uh
278. 10:19stress test the strategy to ensure the
279. 10:21result isn't overfit. So, this is how
280. 10:23the results would look. So, this yellow
281. 10:26line here is the original backtest and
282. 10:29these blue lines here are the
283. 10:30simulations. So, as you can see, the
284. 10:32Sharpe ratio of the original is 1.5. The
285. 10:35median is 1.41 and the best 5% is two.
286. 10:39So, the Sharpe ratio of the original is
287. 10:41clearly way closer to the median number,
288. 10:43which is a really good sign. If this
289. 10:45number was closer to two, such as 1.8,
290. 10:491.9 or even if it was better than two,
291. 10:52so it was clearly in the best 5% that
292. 10:54would have been a clear signal of the
293. 10:56strategy being overfit. But now that it
294. 10:58is not, this is a good sign that it most
295. 11:00likely isn't overfit. And I'm
296. 11:02emphasizing on the word most likely
297. 11:04because you can never be 100% sure with
298. 11:07over fitting, so that's just something
299. 11:08to remember, which is why we want to
300. 11:10find as many as strategies as possible
301. 11:12and run them simultaneously so that if
302. 11:15one of them, for example, isn't
303. 11:16performing well, the other would cover
304. 11:18us. All right, so let's go back. Next,
305. 11:20let's check out this second strategy
306. 11:22that we found for us. So, the backtest
307. 11:24results and the Monte Carlo. Let's open
308. 11:26both of them. Here's the backtest. So,
309. 11:28you see this one also looks pretty good
310. 11:30but during this period, which includes
311. 11:32most of 2023, I don't really like these
312. 11:36results. But here you could say it
313. 11:37crushed everything, so at least it's not
314. 11:40really losing money. So I guess if you
315. 11:41run something like this, at worst it
316. 11:43would be neutral and at best it crushed
317. 11:45it, which is a good sign. The average
318. 11:46trades per month is 15, sharp ratio is
319. 11:49one, the win rate is 36%. It executed
320. 11:52741
321. 11:53trades. So overall it looks good and the
322. 11:56monthly returns also aren't that bad.
323. 11:59So, yeah. This also looks good, but I
324. 12:01would say the first strategy that you
325. 12:03found for us is significant better. So I
326. 12:05actually rather run this one. But if you
327. 12:07check out the Monte Carlo, so this one
328. 12:09also looks really good. Like you see the
329. 12:11original backtest is right in the middle
330. 12:13of the simulations, which is a really
331. 12:15good sign. The sharp is one, which is
332. 12:17below the median. So, if you consider
333. 12:19this fact, this one is actually beating
334. 12:21the first strategy because according to
335. 12:23Monte Carlo, it's even less likely to be
336. 12:26overfit, which is a really good sign.
337. 12:28But still considering this and fact that
338. 12:30it didn't really perform well during
339. 12:322023, I prefer to run the first
340. 12:34strategy. But the first strategy didn't
341. 12:37really perform well during the second
342. 12:38half of 2022, which was a bear market.
343. 12:41In the first half, I would say it was
344. 12:42good. So this is also something to be
345. 12:44aware of. And also in the past few
346. 12:46months, like in this period, it wasn't
347. 12:48doing great, which I'm not really
348. 12:50surprised because the market was in a
349. 12:52really bad phase with all the
350. 12:54uncertainty that was in the market
351. 12:56because of the war and everything. So, I
352. 12:58don't really expect any trend following
353. 13:00strategy to have done well during that
354. 13:02period. But the fact that it did make it
355. 13:04back here also gives me confidence in
356. 13:06running it. Now, let's copy the code for
357. 13:09this strategy and go and check it out.
358. 13:11Cuz I want to see, for example, how did
359. 13:13it do the exit of the strategy? So,
360. 13:16first of all, these are the hyper
361. 13:17parameters that the strategy defined.
362. 13:19These are the DNAs that it defined,
363. 13:21which is basically the result of the
364. 13:23optimization. And here we can see the
365. 13:25Bollinger Bands, the trend EMA, the ATR,
366. 13:28and these are the entry rules of the
367. 13:30strategy, which it says the closing
368. 13:32price must be above the upper band of
369. 13:34the Bollinger Bands, and it should also
370. 13:36be above the current 20 EMA, which is
371. 13:39the moving average, and it's doing the
372. 13:42opposite for short positions. Here's the
373. 13:44position sizing, so the entry price is
374. 13:45going to be the current price, which is
375. 13:47a market order. Here's the stop, and
376. 13:49it's using the ATR indicator, and here's
377. 13:52the position sizing, and here's where it
378. 13:54submit the buy order. So, it's doing the
379. 13:56opposite for short positions, and once
380. 13:57the position is open, using the on open
381. 13:59position event hook, it's saying if it's
382. 14:02a long position, submit the stop loss,
383. 14:04and we want to use the quantity of the
384. 14:06current position, and this is the price,
385. 14:08and then here's where it is updating the
386. 14:12stop loss order. So, this one is
387. 14:14actually the one that is using a
388. 14:15trailing stop. So, not the one that was
389. 14:19supposed to set a specific targets for
390. 14:21closing the position. And if we check
391. 14:23the other one, here we can see again the
392. 14:26Bollinger Bands, trend EMA. So, they're
393. 14:28actually pretty similar, except that in
394. 14:31this one, we aren't submitting a
395. 14:33trailing stop. So, you see, in the
396. 14:35previous one, we were using the update
397. 14:36position method, which is this function
398. 14:38of Jesse, which gets called every time a
399. 14:40new candle closes. So, basically,
400. 14:42whenever we close any candle, we want to
401. 14:45update the stop loss. That's why we call
402. 14:47it a trailing stop. But in this one,
403. 14:49we're not doing that. So, once the
404. 14:50position opens, we submit both the stop
405. 14:52loss and the take profit, and that's it.
406. 14:55We wait until one of them is hit. And
407. 14:57according to this, we're getting better
408. 14:59results if we were trading this. And
409. 15:01actually, guys, this is exactly what I
410. 15:03usually get in my trading, that whenever
411. 15:06I submit specific points for a stop loss
412. 15:08and take profit, I get way better better
413. 15:10results if I were to use a trailing
414. 15:12stop. But a trailing stop is something
415. 15:14that manual traders really like, and
416. 15:16they think it gives them better odds,
417. 15:18but I never saw the same thing in the
418. 15:20algo trading. So, that's one of the
419. 15:22reasons why I like backtesting because
420. 15:24you will be able to debunk some of the
421. 15:25myth of the trading that's out there and
422. 15:28you can see the results for yourself.
423. 15:30All right, so overall I liked the
424. 15:31results that we found, but one thing
425. 15:34isn't exactly how I wanted it. So,
426. 15:35initially when I wrote that prompt, I
427. 15:37expected it to find strategies that can
428. 15:40get me those results on each of these
429. 15:42symbols. I didn't mean it to run them
430. 15:44simultaneously like in this case. So,
431. 15:46this is basically a portfolio of three
432. 15:49symbols. And the entire portfolio is the
433. 15:52one that is giving us these results. So,
434. 15:53if you were to run these individually,
435. 15:56you would not have gotten such a good
436. 15:57number. And I'm not saying that you
437. 15:59shouldn't run multiple symbols
438. 16:00simultaneously. I'm just saying that's
439. 16:02not what I expected from the model.
440. 16:04Probably a misunderstanding on my side.
441. 16:06I should have been more specific with
442. 16:07the prompt. But overall, it did a good
443. 16:09job and it did it very fast. So, all in
444. 16:12all, I would say this is a good model
445. 16:14and I will probably use it, especially
446. 16:16if a speed is something that I care for
447. 16:18because according to the benchmarks,
448. 16:20it's not beating Claude Opus yet. So, if
449. 16:23you want the absolute best quality, this
450. 16:24probably isn't the answer. But if you
451. 16:26need quality and speed at the same time,
452. 16:29this is a better model, especially
453. 16:31because in our use case where we give
454. 16:33tools to the model, so specifically the
455. 16:35MCP tools, it's not just about the
456. 16:37quality of the model. So, the fact that
457. 16:40the model can run faster, do research
458. 16:42faster, that could sometimes cause
459. 16:45better results in the end, which is what
460. 16:47we care about. Now, I am going to submit
461. 16:49this strategy on our strategies page on
462. 16:52our website. So, if you haven't checked
463. 16:53it out, please do so. There are so many
464. 16:55strategies here. You can sort them by
465. 16:57sharp ratio and the period and stuff.
466. 16:59You can also see the details of those
467. 17:02strategy results for other trading
468. 17:04periods, symbols, and time frames. So,
469. 17:06definitely check it out. You can also
470. 17:08copy the source code of these strategies
471. 17:10that I just shared with you guys on our
472. 17:12website. If you enjoyed the video,
473. 17:14please make sure to give it a like and
474. 17:15subscribe to the channel if you haven't
475. 17:17already because I publish tutorials just
476. 17:20like this one all the time. And at the
477. 17:21end, let me know your experience with
478. 17:23Cloud Code in the comments. Thank you so
479. 17:25much for watching, guys. I'll see you in
480. 17:27the next time.