I Let Grok 4.6 Research Crypto Trading Strategies — Transcript

Full transcript

1. 0:00Elon Musk's xAI just launched Grok 4.6
2. 0:04and they are making really bold claims
3. 0:06with this. They're claiming that it is
4. 0:08even beating Claude 5 on the max effort
5. 0:11in some of the benchmarks or at least in
6. 0:13the ones that it isn't, it is at least
7. 0:15beating GPT 5.6 soul, which in these
8. 0:18days is my daily driver and I love that
9. 0:21model because the performance to price
10. 0:23of it is just amazing. So, as you can
11. 0:25see in most of the benchmarks that is
12. 0:27the case. So, they're claiming that it
13. 0:30is beating all other models out there.
14. 0:32And not only that, here's the really
15. 0:34crazy part. It is $2 per million input
16. 0:37token and $6 per million output token.
17. 0:39This is even cheaper than Claude Sonnet.
18. 0:42So, this is just simply crazy. They're
19. 0:44also offering a 2x included usage inside
20. 0:47Grok build and cursor. So, right now
21. 0:49everything is hot. If you want to give
22. 0:51it a try, now's probably the best time.
23. 0:53But you know how it is, I don't just
24. 0:54accept the benchmark. I actually test
25. 0:56the model to see how good it is for
26. 0:59writing a strategy. Use running back
27. 1:00test, optimization, Monte Carlo
28. 1:02simulations to ensure it is not overfit
29. 1:05even. So, we're not going to accept what
30. 1:06they're saying, we're going to test it
31. 1:08today. We're going to be writing an
32. 1:09actual strategy. So, if that's something
33. 1:11you're interested in, stick around and
34. 1:13let's see how it goes. As always, I'll
35. 1:15be using Jesse framework for the algo
36. 1:17trading side of things using Python. It
37. 1:20is very easy to get started with the
38. 1:22solution. If you haven't done it
39. 1:23already, just check out the
40. 1:24documentation. I have also made videos
41. 1:26about this in the past, so you can check
42. 1:28them out. You can of course get started
43. 1:30with Docker as well, which is really
44. 1:31easy. Now, to run Jesse, I I am already
45. 1:34inside my Jesse project. So, all I need
46. 1:36to do is say Jesse run. Now, this is not
47. 1:38going to work because I forgot to enable
48. 1:40the correct conda environment. So, for
49. 1:42that I'm going to say conda activate and
50. 1:45then the name of it, which in my case is
51. 1:48Jesse 3.12. There we go. So, now I'm
52. 1:52going to run Jesse run one more time and
53. 1:54this time it works. So, you see it is
54. 1:56printing out the address for the MCP.
55. 1:59I've already added to my code editor. If
56. 2:01you haven't, you just need to copy this,
57. 2:03and then for the editor side of things,
58. 2:06I'm using the Z editor, which is my
59. 2:07favorite these days. And in order to run
60. 2:10the model, I'm using the Open Router
61. 2:12driver. So, if I simply search for
62. 2:14something such as Grok 4.6, I'm going to
63. 2:17find it here, and that's really great
64. 2:19because it gives me access to basically
65. 2:20all the models out there. In order to
66. 2:22add the MCP, you need to go here and
67. 2:24then add server. It will open this
68. 2:26window, click here, and then add remote
69. 2:29server, and then paste in that address,
70. 2:32and give it a name, and that's it. We're
71. 2:33good to go. Now, I've already done this,
72. 2:35so instead I'm I will just ask, "Hey, do
73. 2:38you have access to Jesse MCP?" So,
74. 2:40always do this because we want to ensure
75. 2:42that it has access to the MCP before
76. 2:44continuing. So, there we go. It says
77. 2:45that it has access to Jesse MCP. Now, by
78. 2:48the way, sometimes it says no, and it's
79. 2:50not just limited to Jesse MCP. It could
80. 2:52happen with any MCP server. And if that
81. 2:55happened to you, first ensure the MCP
82. 2:57server itself is run, or restart it just
83. 2:59in case. So, for example, in my case, I
84. 3:01could just stop this and run the Jesse
85. 3:04run command one more time. And then, you
86. 3:06need to reload your code editor. So, if
87. 3:09I do that, it will probably work. Now,
88. 3:11this was already working for me. So, now
89. 3:13that it has access to it, we can
90. 3:15actually give it our prompt. Or, we
91. 3:16could also create a new session if you
92. 3:18want. Now, the prompt is what we
93. 3:20actually want from the model to do for
94. 3:22us. And I am going to include it in the
95. 3:24source code later. But, for now, let's
96. 3:27just read that out loud. "Hey there,
97. 3:28research and find a trend following
98. 3:30strategy for trading SOL/USDT and
99. 3:33ETH/USDT. It needs to be a squeeze type
100. 3:35of strategy using Bollinger Bands as the
101. 3:37base indicator, along with any other
102. 3:40indicator you find appropriate. You do
103. 3:43not have to use one exact strategy and
104. 3:45set of parameters for both symbols. Feel
105. 3:47free to define two separate strategies
106. 3:49if necessary. However, the entire
107. 3:51portfolio must meet the requested
108. 3:53criteria. I need the results to be good
109. 3:56from the beginning of the year through
110. 3:58now. By good, I mean a sharp ratio of at
111. 4:00least 1.5, risk 3% of the account on
112. 4:03each trade. Every time you develop a
113. 4:05strategy, validate the results using
114. 4:07statistical significance tests before
115. 4:09writing the full strategy. Only proceed
116. 4:12if the strategy's metrics demonstrate
117. 4:14genuine statistical significance. Use
118. 4:16optimization to improve the results. At
119. 4:18the end, apply Monte Carlo simulations
120. 4:20to ensure the results are not overfit.
121. 4:22Continue until you find a strategies
122. 4:24that fully meet the requested criteria.
123. 4:27Do not prompt me in the meanwhile. Good
124. 4:29luck. So, there we go. Let's hit enter,
125. 4:31and we can see it started the
126. 4:33development. It is thinking now because
127. 4:35this is a thinking model, of course. So,
128. 4:36let's go and come back when the results
129. 4:38are ready. So, just a quick update,
130. 4:40guys. I faced a couple of issues with my
131. 4:43open browser setup on Z editor. There
132. 4:45seems to be a bug with the Z editor, and
133. 4:47it couldn't auto compact things. So, I
134. 4:50kept facing errors because of that. And
135. 4:53also, the model just wasn't performing
136. 4:55well at all. I also faced another issue,
137. 4:57and that is my usage was going through
138. 4:59the roof. I mean, look at this chart. I
139. 5:01have spent $52 on Grok 4.6 without even
140. 5:05it getting the job done for me. And this
141. 5:07was supposed to be a cheap model. So, I
142. 5:08did some research, and apparently, using
143. 5:11Cursor is the most affordable way for
144. 5:13using this model. Even their own cloud
145. 5:15code competitor, which is called Grok
146. 5:17Build, is not as good of a deal. So,
147. 5:20people are complaining about the limits
148. 5:22on Reddit. So, I'm going to stay away
149. 5:24from this. So, instead, I'm going to go
150. 5:26with the first plan of Cursor. But, I
151. 5:28don't want to install the Cursor editor.
152. 5:30I just don't like that VS Code fork.
153. 5:34Now, the way I'm going to actually run
154. 5:36Cursor is inside the terminal. So, in my
155. 5:39Z editor, I simply opened a new terminal
156. 5:42agent window. And to run the Cursor CLI,
157. 5:46I have to say agent. There we go. So,
158. 5:48everything's ready and I have already
159. 5:51authenticated. So, let's ensure that we
160. 5:53have the correct model. So, Cursor
161. 5:55Curate 4.6. And then, I want to ensure
162. 5:58that I can run everything so that I
163. 6:00don't have to keep approving the
164. 6:03permissions that the model is asking cuz
165. 6:05that is super annoying. So, I'm going to
166. 6:07click on this and it is on now. So, next
167. 6:10I'm going to paste the entire prompt one
168. 6:12more time and hit enter. So, now it says
169. 6:15it cannot find the Jesse MCP. So, let's
170. 6:18stop it. So, let's hit /mcp.
171. 6:22And yes, we can see that Jesse MCP is
172. 6:24not one of it. All right. So, let's exit
173. 6:26this. And in order to add an MCP, I'm
174. 6:30going to have to edit this file. So,
175. 6:31let's just simply say Z and then paste
176. 6:34in the path for the MCP configuration of
177. 6:36Cursor. There we go. And we can see I
178. 6:39already have one from before. So, let's
179. 6:42just copy this one, paste in a new one.
180. 6:45I'm going to call it Jesse. We do not
181. 6:47need any headers. And let's replace this
182. 6:50with the address of our Jesse instance's
183. 6:52MCP, which is this. Localhost port 9007
184. 6:56and then a /mcp. So, let's save it. And
185. 6:59now, hopefully, if I run the Cursor
186. 7:02agent command one more time. By the way,
187. 7:04it works with both agent or
188. 7:06cursor-agent. So, both of them should
189. 7:08work. Now, let's
190. 7:10run the MCP command. All right. So, we
191. 7:12can see it's been enabled. This is
192. 7:14great. So, let's go ahead and give it
193. 7:16that prompt one last time.
194. 7:20There we go. It can now access
195. 7:22everything. Actually, it's not running
196. 7:24on the run everything mode. So, let's
197. 7:26stop it just one more time. I'm going to
198. 7:28say run everything. All right. So, it is
199. 7:30on now. And let's also create a new
200. 7:33session just in case. Now, I'm going to
201. 7:36give it my prompt. All right. Looks like
202. 7:38finally everything is working. Let's go
203. 7:40and come back a bit later. All right.
204. 7:41So, it finally finished, but before I
205. 7:43show you the results, I want to show you
206. 7:45this that in the usage section of my
207. 7:48cursor account, I can see that it spent
208. 7:512% of my monthly tokens on this. And I
209. 7:54had spent 20 bucks on this. So that
210. 7:56means it was a great deal because now
211. 7:58I'm actually able to continue coding
212. 8:00with this. And besides the usage is
213. 8:03specifically for cursor models, I also
214. 8:04have this one which will give me at
215. 8:06least $20 worth of API usage on other
216. 8:10models. So overall, this was a great
217. 8:12deal, but the one that I was using on
218. 8:14open router was a horrible deal.
219. 8:17Although in the past, I had a pretty
220. 8:19good success using open router. So it's
221. 8:22not like I'm not going to do that again,
222. 8:23but with this specific model, I don't
223. 8:25know why, but this was the case. So I
224. 8:27just wanted you guys to know this. So
225. 8:29it's finished the results and for the
226. 8:32ETH, it was able to find a strategy with
227. 8:34a sharp ratio above 1.5. It was passing
228. 8:36the RST which is stands for rule
229. 8:38significance testing which is this test
230. 8:41in the just framework which lets us know
231. 8:44whether the integer rules of the
232. 8:46strategy were based on luck or did it
233. 8:49have a genuine edge. It also passed the
234. 8:52Monte Carlo test. So this is good and
235. 8:54here we can see the URLs for the RST. So
236. 8:57let's take a look. I'm going to make
237. 8:59this a little bit bigger so I can see
238. 9:01the URLs in full length. So let's open
239. 9:04the RST for the easy strategy and the
240. 9:07result of its backtest and the result of
241. 9:09its Monte Carlo test. Now while that's
242. 9:11going, I want to quickly remind you guys
243. 9:13about our Telegram. It's the fastest way
244. 9:15to get notified about my future work
245. 9:17whether it's a new tutorial or a tool
246. 9:19that I create. Also, don't forget to
247. 9:20check out our free Discord where more
248. 9:22than 5,000 members like you and I are
249. 9:24hanging out there and helping out each
250. 9:25other with algo trading so we can all
251. 9:27succeed together. The links for both are
252. 9:29down in the description. So there we go.
253. 9:31This is the RST and as you can see the
254. 9:33return of the strategy is by far beating
255. 9:36the random simulations, which is a good
256. 9:39sign. We can also read it here that the
257. 9:41P value is below the threshold, and
258. 9:44hence it does have an edge. The analyzed
259. 9:46return is also 13%, so this is good. And
260. 9:49here we can see the backtest results,
261. 9:52and the equity curve looks fine. I
262. 9:54wouldn't say it's awesome, but it's not
263. 9:55that bad. But this is just for one
264. 9:57strategy, and we were supposed to run
265. 9:59two strategies at the same time. So,
266. 10:01this isn't that bad. This drawdown chart
267. 10:04also doesn't look great. So, over these
268. 10:06months, I don't like how it looks. The
269. 10:08monthly return isn't that bad. It's
270. 10:10positive almost on every single month.
271. 10:12So, overall, looks okay. The P&L was
272. 10:1516%, the minute was 51%, the Sharpe
273. 10:18ratio is 1.69. So, overall, I would say
274. 10:21pretty okay. Let's take a look at the
275. 10:22Monte Carlo. So, with this test, we can
276. 10:25see the Sharpe ratio of the original
277. 10:27backtest is 1.56,
278. 10:29which is above the best 5%, which was
279. 10:321.61.
280. 10:33So, this is a good sign that the
281. 10:35strategy isn't overfit. However, take
282. 10:37this with a grain of salt because even
283. 10:40though I usually want the Sharpe ratio
284. 10:42to be below this, and it is in this
285. 10:44case, but the median itself is negative.
286. 10:47Also, yes, I want the Sharpe of the
287. 10:49original backtest to be below this
288. 10:50number, but I also want it to be like
289. 10:52right in the middle and close to this
290. 10:54value. But right now, we can see it that
291. 10:57is not the case. In fact, this value is
292. 10:58negative. So, overall, I wouldn't call
293. 11:01this Monte Carlo a good result, and I
294. 11:04personally wouldn't use this strategy
295. 11:06just yet. But let's move on because we
296. 11:08have another strategy for SOL. So, here
297. 11:11we can see again it is passing the RST.
298. 11:14The Sharpe was 2.2. This is good. The
299. 11:16return was 18%. Backtest results should
300. 11:19look good on this. And there we go.
301. 11:21Here's the result. So, I wouldn't say
302. 11:23this is super great, but definitely not
303. 11:25bad. The drawdown chart actually looks
304. 11:27much much better. So, I'm happy with
305. 11:29this so far. The Sharpe was two, and the
306. 11:33piano was 18%. The max drawdown is minus
307. 11:364%. So this is awesome actually, because
308. 11:38I can easily add to the size of the
309. 11:39position. So overall, this looks really
310. 11:42good. Next, we have the Monte Carlo
311. 11:44result, which even the model itself is
312. 11:46saying that it is overfit, because the
313. 11:47original backtest had a sharp of 1.83,
314. 11:51which is clearly above 1.18. So let's
315. 11:54open it, because I just want to show you
316. 11:55how the result of a overfit strategy
317. 11:58looks like. And there we go. So you see,
318. 12:00the original backtest is clearly beating
319. 12:03the simulations. And by the way, this is
320. 12:05basically a stress test for the
321. 12:07strategy. So it's not just like the rule
322. 12:09setting quits testing, where we actually
323. 12:12wanted the original backtest to beat the
324. 12:14simulation. So in this case, we don't
325. 12:16want that to be the case. You want it to
326. 12:17perform well even during the
327. 12:18simulations. So let's go back, so we can
328. 12:20check out the result for both the
329. 12:23strategies being run at the same time.
330. 12:24In other words, the entire portfolio. So
331. 12:27let's click here. There we go. So this
332. 12:29actually looks really, really awesome.
333. 12:31So this equity curve looks significantly
334. 12:33better. If you take a look at the
335. 12:35drawdown chart, we can see that the
336. 12:37equity curve is actually pretty good.
337. 12:39Like during this period, it was awesome.
338. 12:42We had a period where the equity curve
339. 12:44was basically in a range, and normally I
340. 12:47would accept this, because at least it's
341. 12:49not going down, and in other parts it is
342. 12:51going up. Also, I should mention that
343. 12:54the whole period that we gave this
344. 12:55research, which is basically since the
345. 12:57beginning of the year until now, isn't
346. 12:59really long enough to find a strategy
347. 13:01that we have a good confidence in.
348. 13:04Usually, if you do have the data, you
349. 13:06want to spend more time on this. But
350. 13:07anyway, so the monthly return also looks
351. 13:11great. Like we just have one month with
352. 13:13a minus 0.7%
353. 13:15return. Other months, we are in a
354. 13:17profit. So that's great. And if we take
355. 13:20a look at the metrics, so it executed 74
356. 13:23trades. The return was 38%. The max
357. 13:26drawdown was minus 4.5. So, this means I
358. 13:29can easily multiply the position sizing
359. 13:32of the strategy and by doing that the
360. 13:34return is going to go way above 100%
361. 13:37because usually I'm comfortable up to
362. 13:40minus 30% max drawdown. So, that means I
363. 13:43could even six times the position sizing
364. 13:45and by doing that this is going to be
365. 13:47about 240%
366. 13:49return. The strategy was paying trading
367. 13:51fees. The win rate was 48% with an
368. 13:55average win to loss ratio of 2.22. So,
369. 13:57this is really good. The sharp ratio was
370. 13:592.68 and the average trade per month was
371. 14:0210. So, overall pretty good results.
372. 14:04Let's take a look at the trade chart.
373. 14:06So, this chart will show us the details
374. 14:09of what were the indicator values when
375. 14:12it executed trade. Now, let's look at
376. 14:14the candles first. So, here we can see
377. 14:16we are showing the values for the
378. 14:19Bollinger Band upper band, the mid, the
379. 14:21lower band and we also have the Keltner
380. 14:24Channel which is this orange line here,
381. 14:25but it's a bit too heavy I would say.
382. 14:27So, let's toggle them. All right, so now
383. 14:30for example we are only displaying the
384. 14:32Keltner Channel and this is much easier
385. 14:35to read. We also have other charts. So,
386. 14:37the squeeze basically. So, whenever it
387. 14:39is one we are ready to take a position
388. 14:42like in this case for example and
389. 14:44whenever it is zero we're not going to
390. 14:46do that. So, this is super helpful
391. 14:47because it allows us to go deep to
392. 14:49ensure the strategy is performing the
393. 14:51way that we expect it to. We also have
394. 14:53these trade history here. So, if I click
395. 14:56here for example, I can see exactly this
396. 14:58trade like when did it open, when did it
397. 15:01close. Let's go check out another one.
398. 15:03So, this was a profitable trade. So, you
399. 15:05can see it opened it right here when the
400. 15:07breakout happened. You know, or in other
401. 15:10words when a squeeze actually happened
402. 15:12and then we closed it here. We can also
403. 15:13switch this to check out the trades for
404. 15:16the other strategy. Again, we can toggle
405. 15:18these to see the values that are helpful
406. 15:21to us. We also have different charts
407. 15:23here. So, this one, for example, is
408. 15:25displaying the slope, which is a pretty
409. 15:28useful indicator. So, overall, this
410. 15:30backtest chart actually looks really
411. 15:32great. So, now, let's check out the
412. 15:35Monte Carlo results for it. There we go.
413. 15:37So, the original backtest result is
414. 15:40beating the simulations like almost all
415. 15:42of them. This is a huge red flag. And if
416. 15:45you read it here, the Sharpe ratio is
417. 15:472.70, which is way above even the best
418. 15:505% of the Monte Carlo simulations. So,
419. 15:53this is like a textbook overfit
420. 15:56strategy, which means even though on the
421. 15:58backtest we were getting these good
422. 16:00results, we cannot really trust it yet.
423. 16:03Now, by the way, it doesn't mean that
424. 16:04the strategy is definitely helpless and
425. 16:07that we cannot improve it or that it
426. 16:09will definitely fail, but it also means
427. 16:12that we do not have the safety of
428. 16:14passing a Monte Carlo test before moving
429. 16:16into live trading. So, I personally
430. 16:18would not be trading a strategy like
431. 16:20this at all. So, this is a really good
432. 16:22example of how an overfit strategy looks
433. 16:25like. So, we usually show the ones that
434. 16:27look good. So, in this video, I'm
435. 16:29showing you one that actually isn't.
436. 16:31Now, I am going to add this to our
437. 16:33strategies index page, which is where I
438. 16:36submit the strategies that I develop.
439. 16:39It's mostly me, but sometimes other
440. 16:40users also do this. You can also check
441. 16:42out more details about each strategy in
442. 16:44different periods, symbols, or time
443. 16:47frames. So, this is also super helpful.
444. 16:48And again, I'm going to submit that
445. 16:50strategy as well. So, if you're curious
446. 16:52to know its results in other situations,
447. 16:54definitely check out this page, and I am
448. 16:56going to link to it in the description.
449. 16:58It is also the best way to get the
450. 16:59source code of the strategy that I just
451. 17:01showed you guys. So, if you want to run
452. 17:03it yourself, feel free to do that. So,
453. 17:05overall, my experience with this model
454. 17:07just wasn't good because even though on
455. 17:10the benchmarks it is showing that it is
456. 17:12beating GPT 5.6, and in some of them it
457. 17:16is even beating favorable 5 model.
458. 17:18That's not what I saw. For example, if
459. 17:20you check out some of my previous videos
460. 17:22such as this one, which I used the
461. 17:24Claude Opus 5 model or Kimmy K3, they
462. 17:27were easily meeting the criteria of the
463. 17:30research that I asked the model to do
464. 17:33even on the same kind of squeeze
465. 17:35strategy for ETH. But not only this one
466. 17:37failed, it also it stopped very early.
467. 17:40So, you could argue that maybe I should
468. 17:41have continued and asked the model to
469. 17:43keep going, but the fact that it stopped
470. 17:45early, that was a red flag as well.
471. 17:48Also, this wasn't even my first attempt.
472. 17:50You remember that I tried it twice in
473. 17:52this video. So, first time I used the
474. 17:54agent panel of the Z editor, and while
475. 17:57that one was failing and burning a lot
476. 17:59of tokens for me, I switched into the
477. 18:01cursor agent. And while this one went
478. 18:03through, it couldn't meet the criteria.
479. 18:05Also, at the pricing seems really low,
480. 18:07like it is $2 per million input token
481. 18:11and $6 per million output token, which
482. 18:14is significantly low. It's even lower
483. 18:16than the Claude Sonnet model, but still,
484. 18:19you remember that it very easily burned
485. 18:21$52 with just one experiment. So, I
486. 18:24don't know what's going on with this
487. 18:25release, and I've seen some people
488. 18:27claiming that the model is good,
489. 18:29especially on YouTube, but that wasn't
490. 18:31my experience. So, anyways, I hope you
491. 18:33guys enjoyed the video. If you did,
492. 18:34please give it a like and post a comment
493. 18:36so let me know what your experience with
494. 18:37this model if you have already tested it
495. 18:39so far, and don't forget to subscribe
496. 18:41for future videos just like this one.
497. 18:43Thank you so much for watching. I'll see
498. 18:45you in the next one.