Grok 4.5 + MCP finds a 1.9 Sharpe Strategy in 11 minutes! — Transcript

Full transcript

1. 0:00Elon Musk's SpaceX
2. 0:02Wait, that doesn't sound right. Just
3. 0:04released Grok 4.5. And not only in some
4. 0:07of the benchmarks, it is reaching to
5. 0:09Claude Fable 5, but it is significantly
6. 0:12cheaper. So, in other words, the quality
7. 0:15to price ratio of this model is better
8. 0:18than any other model out there. And that
9. 0:20makes it a great candidate for us to do
10. 0:23algo trading research with it. As
11. 0:25always, what I'm curious about is
12. 0:27whether this model can write me trading
13. 0:29strategies, run back test optimization,
14. 0:32and also Monte Carlo in order to make
15. 0:34sure that the strategies that it is
16. 0:36developing for me are not overfit. So,
17. 0:38that's exactly what I'm going to do in
18. 0:39this video. I'm going to test the model
19. 0:41and ask it to write me real-world
20. 0:42strategies. Before we continue, I got to
21. 0:44mention I am not a financial advisor,
22. 0:46and this video is for educational
23. 0:48purposes only. So, with that out of the
24. 0:50way, let's get right to the video.
25. 0:53Now, here's the blog post released by
26. 0:56the Grok team, and here are the
27. 0:57benchmarks. So, you see in this one, it
28. 1:00is right behind GPT-5. In this one, in
29. 1:02this one, it is right before Opus 4.8.
30. 1:06In this one, it is number one, so it's
31. 1:08even beating Fable. And in this one, it
32. 1:10is very close to Fable and GPT-5.5. But
33. 1:14in this one, Fable is beating everything
34. 1:16out of the water. Now, one thing we have
35. 1:19to consider is that not only the price
36. 1:21of this model is significantly cheaper,
37. 1:24but it is also significantly faster in
38. 1:26execution. And it is apparently also
39. 1:29more efficient in token usage. So, that
40. 1:31means that not only you'll you'll be
41. 1:33getting a good enough quality, but you
42. 1:35will also get it at a very much faster
43. 1:38face. Now, a little bit of spoiling of
44. 1:40the video, that's exactly what I got at
45. 1:42the end of it. So, the result that I
46. 1:44got, it was significantly faster than
47. 1:47any other model that I have seen. But
48. 1:48anyways, I don't want to expose just
49. 1:50everything. So, let's continue. Here, it
50. 1:52is claiming that this model is faster
51. 1:54than flash models, and the speed is 80
52. 1:57TPS or token per second, which is
53. 1:59absolutely amazing. And you see it is
54. 2:02also 4.2 times more efficient than Opus
55. 2:064.8 in token usage, which is of course
56. 2:09amazing. Now, for the algo trading side
57. 2:12of things, as always, I'll be using the
58. 2:13just framework, which also gives us an
59. 2:16MCP, which I can just connect it to my
60. 2:19agent
61. 2:23the work. And for the editor, I'm going
62. 2:26to be using the Zed editor, which is my
63. 2:28new favorite code editor. It is written
64. 2:30in Rust. It is extremely fast and just
65. 2:32amazing, and it is also free to get
66. 2:34started with. All right. Now, as for
67. 2:36about the prompt, I've prepared this,
68. 2:39which by the way, I will open source
69. 2:41this and put it on GitHub and link to it
70. 2:44in the description of the video in case
71. 2:45you want to copy and paste it. But, I
72. 2:47will read it very fast because this is
73. 2:49the most important part of the video.
74. 2:51Hey there. Do research and find me three
75. 2:53trading strategies. Each one of them
76. 2:55should have a sharp ratio above one
77. 2:57since the beginning of the year until
78. 2:59beginning of July. They should be trend
79. 3:01following, include both and short
80. 3:03positions filtered by a market regime
81. 3:05indicator, such as the ADX, so trades
82. 3:08only fire in trending conditions, and
83. 3:10risk 3% of the account capital per each
84. 3:13trade. However, since we are in a bear
85. 3:15market, I want them to have more focus
86. 3:18on short trades. And the entry rules for
87. 3:21short and long positions should not
88. 3:23necessarily be the same. In the first
89. 3:25strategy, once the position is open,
90. 3:27close it using a trailing stop based on
91. 3:30some sort of indicator. In the second
92. 3:32strategy, close the position using a
93. 3:35fixed multiple of ATR from entry. You
94. 3:38can test them on BTC, ETH, and SOL,
95. 3:41Trump, Melania, and DOGE on the hourly
96. 3:45and 4-hour time frames. The strategies
97. 3:48should trade these symbols individually
98. 3:50or altogether. Every time you develop a
99. 3:52strategy, validate the results using a
100. 3:55statistical significance test before
101. 3:57writing the full strategy. Only proceed
102. 3:59if the strategy's metrics demonstrate
103. 4:01genuine statistical significance. Feel
104. 4:04free to use optimization to improve the
105. 4:06results. At the end, apply Monte Carlo
106. 4:09simulations to ensure the results are
107. 4:10not overfit. Continue until you find
108. 4:13strategies that fully meet the criteria
109. 4:15requested. Do not prompt me in the
110. 4:17meanwhile. Good luck. And by the way,
111. 4:19guys, I am using the Z editor, and the
112. 4:21way that I am using Grok 4.5 is by
113. 4:26picking it from here. So, I have already
114. 4:28added my Open Router credentials, and
115. 4:31now I I am able to pick whichever model
116. 4:34that's being offered by them. And if I
117. 4:36simply look it up, I can easily find
118. 4:38Grok 4.5 and pick it. Now, the catch is
119. 4:41that at the moment of recording the
120. 4:42video, Grok 4.5 is not available to EU
121. 4:46citizens, and the way I am accessing it
122. 4:48is by simply using a VPN. Because if I
123. 4:51were to use a an EU IP, this would have
124. 4:54gave me an error. And by the way, here
125. 4:56there is a typo, so let's fix it. All
126. 4:58right. So, everything looks good. Let's
127. 5:00just hit enter and wait for the agent to
128. 5:03do the research. However, there is one
129. 5:05thing that I would always like to do.
130. 5:07So, I'm just going to copy this, delete
131. 5:09it, and first I will say, "Hey, do you
132. 5:11have access to Jesse MCP? Can you do
133. 5:13research for me?" and hit enter. So, the
134. 5:16reason I'm doing this because I want to
135. 5:17ensure that my setup is working before I
136. 5:20give the agent some sort of task that is
137. 5:22really huge. Because if it doesn't have
138. 5:25access to the correct tool, then it
139. 5:27would be pointless for me. All right, so
140. 5:28it turns out my VPN wasn't working, so
141. 5:30I'm just going to fix that. All right,
142. 5:32so it seems to be figured out, so let's
143. 5:34retry this, and this time hopefully it
144. 5:37should work. And there we go. So, yes,
145. 5:39it has access to Jesse MCP. It says that
146. 5:41everything is fine. All right, so let's
147. 5:43begin the actual research process. So,
148. 5:45I'm going to paste back to our prompt
149. 5:48and hit enter. All right. So, as you can
150. 5:50see, it is using the MC tool just fine.
151. 5:54It is listing the indicators that it has
152. 5:56access to. It's getting details because
153. 5:58it is preparing to start writing
154. 6:00strategies. And it is a thinking model,
155. 6:02of course, so it's going to do lots of
156. 6:04thinking. And here also, if I hover my
157. 6:07mouse on it, I can see the context. So,
158. 6:1012% has been already used and this model
159. 6:13has 500 thousand context window, which
160. 6:16is pretty big. So, it's not 1 million
161. 6:18like Opus, that is how I would like it
162. 6:21to be, but it's still 500 thousand is a
163. 6:24lot. So, hopefully it's going to be more
164. 6:26than enough for our use case. All right.
165. 6:28So, I'm going to go back and come back
166. 6:30later until this one is updated. All
167. 6:32right. So, I faced one issue and that
168. 6:36was because the limit for my API key had
169. 6:39reached. So, this isn't really something
170. 6:41you should worry about. Just ensure that
171. 6:43the API key that you generate on
172. 6:45OpenRouter has enough credit allowance.
173. 6:47However, next I faced another issue and
174. 6:50I cannot actually see it here to show it
175. 6:52to you, but it was about the context
176. 6:54window. And looking at here, we can see
177. 6:56that 75% of the context window is full
178. 6:59and apparently the request that it tries
179. 7:01to send it is passing this number and as
180. 7:04a result OpenRouter is returning an
181. 7:06error. Now, the thing is the Z editor
182. 7:08recently added an option for auto
183. 7:10compacting, which is really helpful here
184. 7:12because if it compacts the values here,
185. 7:16then we won't have this issue. But I
186. 7:17think the problem is this that the
187. 7:19threshold is set to 90%. So, if we lower
188. 7:22this to, let's say, 70% and hit retry,
189. 7:26hopefully it should be resolved because
190. 7:28now it tries to compact the context
191. 7:30first and then continue everything by
192. 7:32itself. So, apparently it's getting
193. 7:33errors even for compacting stuff. So,
194. 7:35I'm guessing that if we start over, it's
195. 7:38going to work this time. So, I will just
196. 7:40go up here and press enter again. So,
197. 7:43now the context window is cleared again
198. 7:45and this time we have the auto
199. 7:47compacting option set to 70%. So,
200. 7:50hopefully we won't face the same issue
201. 7:51again. Now guys, I could have deleted
202. 7:53this part from the video, but I didn't
203. 7:55because I think this is some sort of
204. 7:57issue that you might also face. So, I
205. 7:59wanted to include that in the video. So,
206. 8:01you guys know how to deal with it if you
207. 8:03face the same issue. And you might be
208. 8:05wondering, "Okay, so why didn't you face
209. 8:07this issue before?" Well, in the
210. 8:09previous videos I was using tools such
211. 8:11as Cloud Code or OpenAI Codex and with
212. 8:14those tools, they handle the auto
213. 8:17compacting by themselves. So, I never
214. 8:19had to worry about it. Sure, I was still
215. 8:21using my Z editor, but those tools were
216. 8:25the ones actually handling the auto
217. 8:27compacting and context management stuff.
218. 8:30But, now that I'm using the Z editor's
219. 8:33built-in agent window, I have to worry
220. 8:36about this. But, thankfully recently
221. 8:38they added that option which allows us
222. 8:40to auto compact. So, I just needed to
223. 8:42change its default threshold from 90
224. 8:45into 70. All right, so the results are
225. 8:48in and it actually took only 11 minutes.
226. 8:51Now, it is removed from here probably
227. 8:53because I gave it another prompt after
228. 8:56this. So, you're just going to have to
229. 8:57trust my word, but it only took 11
230. 9:00minutes for it to meet the criteria that
231. 9:02I had set for the model. That means
232. 9:05three strategies with a sharp ratio
233. 9:07above one. And here's the output. It
234. 9:09gave me couple of tables and URLs for me
235. 9:12to check out the results on the
236. 9:14dashboard with some pretty charts and
237. 9:16stuff. However, I found it to be just a
238. 9:18little bit all over the place because it
239. 9:20gave me like multiple tables and that
240. 9:22just wasn't fine. So, I said, "You know
241. 9:24what? Give me just one table with all
242. 9:27the results combined." And it gave me
243. 9:29this which is perfect. But, before I
244. 9:30open it, I want to show you this that it
245. 9:32also gave me report files and these are
246. 9:35really great. So, for example, let's
247. 9:37open this one and you see it's in the
248. 9:39markdown format. So, if I press command,
249. 9:42shift, and V on my macOS machine, I can
250. 9:45see the rendered result of the markdown
251. 9:48format on my Z editor. And here you can
252. 9:50see that it's telling me this was the
253. 9:53objective, this is the target, the
254. 9:55exchange that it used, how much to risk
255. 9:57per each trade, and the style of the
256. 9:59strategy, and things like that, and a
257. 10:01pretty summary of it. Then we can see
258. 10:03the entry validation or the RST, which
259. 10:05is stands for rule significance test,
260. 10:08which basically tells us if the entry
261. 10:10rules of the strategy were just pure
262. 10:12luck or if there was some sort of edge
263. 10:14in it. And you can see the P value is
264. 10:17really low, which is a great sign that
265. 10:19the entry rules of the strategy were
266. 10:21definitely not noise and there was some
267. 10:23sort of edge in them. And it also give
268. 10:25me the URL for the dashboard. So, if I
269. 10:27wanted to open it in the dashboard, I
270. 10:29could have just click on this. And then
271. 10:31we can see that it's telling me about
272. 10:33the iterations that it took. So, what
273. 10:35was the first version of the strategy
274. 10:37that it wrote, what was the second one,
275. 10:38and what was the final one. And here's
276. 10:41the results of the final strategy. So,
277. 10:44the max drawdown is minus 26%, the net
278. 10:46profit is 83%, the sharp is 1.93, which
279. 10:50is absolutely great, the number of
280. 10:51trades that it took, and things like
281. 10:53that. Next, we have the result of the
282. 10:55Monte Carlo simulation, which is
283. 10:57basically a stress test, which tells us
284. 11:00whether or not the strategy is going to
285. 11:02be overfit or not. Or to put it better
286. 11:05in a more statistical term, it will tell
287. 11:07us how likely the strategy is to be
288. 11:09overfit because you can never be 100%
289. 11:12sure if the strategy is going to be
290. 11:14overfit or not. It's just an estimation
291. 11:16and this is exactly what gives us to us.
292. 11:19And this is also the result. It just
293. 11:21doesn't have the chart. Like otherwise,
294. 11:22I didn't even need to open the Monte
295. 11:25Carlo page. So, anyways, next we can see
296. 11:27that it's telling me, "Okay, did it
297. 11:29reach the target?" Yes, it did. And then
298. 11:31the recommended next step. And we can do
299. 11:33this for all the other ones. So, we have
300. 11:35two more here. Next, let's take a look
301. 11:37at the table that it gave me, so I can
302. 11:40just open the results on the dashboard
303. 11:42so we can see it. So, the backtest of
304. 11:44the first one, the RSC, and the Monte
305. 11:46Carlo. Now, I want to begin with the
306. 11:48significance rule test because this one
307. 11:50is the first priority when you want to
308. 11:53develop a strategy because if the entry
309. 11:55rule of the strategy is noise, and it
310. 11:58doesn't have an actual edge in it, then
311. 12:00all the other results that you're going
312. 12:01to get for it will be meaningless. This
313. 12:03is really important to always begin
314. 12:05with. And in our case, you can see that
315. 12:07yes, this would have been the return of
316. 12:09the strategy in the simulation, and
317. 12:11these are the simulations that were pure
318. 12:14noise. And the fact that the strategy's
319. 12:16returns are significantly higher than
320. 12:18these is a great great sign. And here
321. 12:20you can also see the results translation
322. 12:23from the dashboard. It's telling us that
323. 12:24it is exceptionally strong evidence of
324. 12:28genuine edge. And here we can see the DP
325. 12:30value. Here is the annualized return,
326. 12:33the observed mean, and then we can see
327. 12:35the number of simulations that were
328. 12:37executed, which is 2,000. All right, so
329. 12:39this looks awesome. Let's close it. Now,
330. 12:41while that's going, I want to quickly
331. 12:42remind you guys about our Telegram. It's
332. 12:44the fastest way to get notified about my
333. 12:46future work, whether it's a new tutorial
334. 12:49or a tool that I create. Also, don't
335. 12:50forget to check out our free Discord
336. 12:52where more than 5,000 members like you
337. 12:54and I are hanging out there and helping
338. 12:56out each other with algo trading so we
339. 12:57can all succeed together. The links for
340. 12:59both are down in the description. Next,
341. 13:01we have the result of the backtest
342. 13:03itself. Now, here's the equity curve,
343. 13:05and here's a log version of it. So, we
344. 13:07can see while all the other assets, so
345. 13:09BTC, ETH, and SOL, and DOGE, while they
346. 13:12were all going down, we were making
347. 13:14money here. However, one thing I noticed
348. 13:16is that the strategy only makes money if
349. 13:19there's a huge crash. So, for example,
350. 13:21in this case. You see, but if the price
351. 13:24is in a range, it will slightly bleed.
352. 13:26So, this is something to be aware of,
353. 13:28but then again here we had a huge crash
354. 13:30and so the the strategy's equity curve
355. 13:32just started going up again. So, if
356. 13:34you're going to trade a strategy like
357. 13:36this, you have to be prepared for a low
358. 13:39win rate and the fact that for months it
359. 13:42wasn't doing great because the market
360. 13:44was in a range. So, this shouldn't be
361. 13:46the only strategy that you're trading or
362. 13:48otherwise it's going to be super hard
363. 13:49for you mentally to just wait and let
364. 13:52the strategy do its thing. Next, we have
365. 13:54the five worst max drawdown periods.
366. 13:56This one is the obvious one of course.
367. 13:58Next, we can see the monthly returns.
368. 14:00So, we have three months in green and
369. 14:03three months in negative, but the months
370. 14:05in green are significantly higher, which
371. 14:07is why the annual return is also really
372. 14:09high. Next, we can also take a look at
373. 14:11the metrics of the strategy. So, the P&L
374. 14:14was 82% and by the way, guys, this is
375. 14:17not for the whole year. So, we need to
376. 14:19also remember that. The max drawdown is
377. 14:21minus 26%. The average win to loss ratio
378. 14:24is three, which is a really important
379. 14:26metric for us. The annual return is
380. 14:29240%.
381. 14:30The max underwater period is 116 days.
382. 14:34Again, what I just explained with the
383. 14:36psychology side of the strategy. So,
384. 14:37this one's going to be really hard to
385. 14:39execute mentally. The win rate is 41%,
386. 14:43which basically means only four out of
387. 14:45every 10 trades that the strategy was
388. 14:47taking was in profit and the sharp ratio
389. 14:49is 1.93. And the average trades that we
390. 14:52took per month was 8. 58. So, this is
391. 14:56also pretty important. So, do not expect
392. 14:58a strategy like this to take trades like
393. 15:00every day. Next, we can see the result
394. 15:03of the Monte Carlo simulation, which
395. 15:05will basically tell us how likely the
396. 15:07result of the strategy is to be overfit.
397. 15:09So, you see the original sharp ratio was
398. 15:12two, the median was 1. almost four, and
399. 15:15the best 5% of the simulations was 3.66.
400. 15:19So, the strategy is not likely to be
401. 15:21overfit. It is significantly lower than
402. 15:24the best 5% and pretty close to the
403. 15:26median. So, this is a really good result
404. 15:29and the number of scenarios that we took
405. 15:30in this simulation was 200. So, some
406. 15:33people say increase this to like 500 to
407. 15:361,000. You could do that, but in my
408. 15:38personal test, even 100 was enough for
409. 15:40most cases, but 200 is definitely more
410. 15:43than enough for me. Next, let's take a
411. 15:45quick look at the backtesting and the
412. 15:47Monte Carlo of the other two strategies.
413. 15:50So, this is the result of the
414. 15:51backtesting of this one. Well, sure, it
415. 15:54is making more money during the
416. 15:56downturns, but in the ranging markets,
417. 15:59this one is bleeding significantly
418. 16:00higher than the first strategy. So, this
419. 16:03is not really a good sign and I don't
420. 16:05really think I would want to trade this
421. 16:07one at all. So, let's just jump into the
422. 16:10third strategy. Here's the result of its
423. 16:12backtest. So, again, I would say it is
424. 16:14bleeding more during the ranging market
425. 16:17and this is not really a good sign, but
426. 16:19overall, the result isn't that bad and
427. 16:21if you take a look at the Monte Carlo
428. 16:23result of it, you can see the original
429. 16:25backtest had a sharp ratio of 1.28
430. 16:29while the median was 1.56 and the best
431. 16:325% was 3.40. So, the strategy is not
432. 16:35likely to be overfit or in better terms,
433. 16:38it is less likely to be overfit and I
434. 16:40like these results, but again, I don't
435. 16:42think this one is holding a candle
436. 16:45compared to the first result that we
437. 16:46got. So, all in all, I would say this is
438. 16:48really good result. However, before you
439. 16:51go and jump and start trading this
440. 16:53strategy or something like this,
441. 16:55remember that you should backtest it on
442. 16:57out of sample. So, in this case, we were
443. 16:59using only this year's result, but
444. 17:01before actually starting to use it, you
445. 17:03need to run it on previous year's data,
446. 17:05but that's also going to be some sort of
447. 17:07personal judgement because the thing is
448. 17:09this year's market was very different
449. 17:12than previous years that I've seen. It
450. 17:14was definitely bear market, but not an
451. 17:16obvious one. It's not like we had a real
452. 17:18bear market followed by a bear market.
453. 17:20So, that's not what we got. So, it was
454. 17:22very different. And the reason that I
455. 17:25asked the agent to generate strategies
456. 17:27that were more focused on shorting was
457. 17:30because I believe that we are still in a
458. 17:31bear market. But if my belief is wrong,
459. 17:34there's a good chance that these
460. 17:35strategies are going to stop working.
461. 17:37So, I might also run a back test for it
462. 17:39for, let's say, 2022 because it was a
463. 17:42similar year. But overall, in algo
464. 17:44trading you want to back test your
465. 17:46strategy on as much data as possible to
466. 17:49gain more confidence in it before
467. 17:51actually risking your hard-earned money.
468. 17:54Now, in case you are curious to know how
469. 17:56much tokens I just burned for this
470. 17:58experiment using the Grok 4.5 model,
471. 18:01well, here's the result. I spent about
472. 18:04$18 on it, more than 13 million tokens
473. 18:07in total with a cash hit rate of 81%.
474. 18:10Now, for $18, which is almost 20, you
475. 18:13could definitely get a subscription plan
476. 18:15of a provider such as Anthropic, OpenAI,
477. 18:19GLM, you name it. So, there are so many
478. 18:21options out there. And if you're going
479. 18:22to spend this much, I don't think it's
480. 18:24feasible to spend it on API tokens. Now,
481. 18:28the Grok team or xAI is offering their
482. 18:31own coding plans, apparently, and they
483. 18:33begin from $30. But the CLI, which I'm
484. 18:36guessing is the one that we need, is
485. 18:38called Super Grok Heavy. So, the light
486. 18:41or this plan is not going to be enough
487. 18:43for us for coding. And the Grok Heavy
488. 18:46begins from $100 per month, and this is
489. 18:49just for 3 months. So, after that, it's
490. 18:50going to be $300 per month. Now, for the
491. 18:53amount of tokens that they are offering,
492. 18:56this might be good for you, but I
493. 18:57personally do not find this a good
494. 18:59enough deal because on Anthropic, for
495. 19:01example, you can get the starting max
496. 19:04plan for just $100, and it will be more
497. 19:08than enough. So, I don't think the Grok
498. 19:10team is offering a good deal at the
499. 19:12moment. So, even though that the
500. 19:13performance of the model was really
501. 19:15good, this was just over the API. And
502. 19:17sure, apple-to-apple comparison, like if
503. 19:19you were to compare the Grok API to
504. 19:22another API such as, let's say, Claude
505. 19:25Opus or Fable, yes, this this Grok model
506. 19:28would have been a better deal. But now
507. 19:30that we have subscription plans on other
508. 19:31providers, I don't think it is exactly
509. 19:34what I would want to use. So, just be
510. 19:36careful about that. And as always, I
511. 19:38will be submitting these strategies that
512. 19:40I just showed you guys to our strategy
513. 19:42index page, where you can browse
514. 19:44previously submitted strategies by me or
515. 19:46others. And you can also click on a
516. 19:48certain strategy and get to see its
517. 19:51results for different trading periods,
518. 19:53symbols, or time frames, which is pretty
519. 19:54helpful. You can also get the source
520. 19:56code of the strategies by clicking here,
521. 19:58of course. I hope you guys enjoyed the
522. 19:59video. If you did, please like the video
523. 20:01and let me know what you think in the
524. 20:03comment sections. And make sure to
525. 20:05subscribe to the channel because I
526. 20:06release tutorials just like this one all
527. 20:08the time. Thank you so much for
528. 20:10watching. I'll see you in the next one.