GPT OSS tested: Is it good for algo trading strategies? (+ fully local installation) — Transcript

Full transcript

1. 0:00Hey guys, it's S. Open AAI just released
2. 0:02two open-source models in the sizes of
3. 0:0520 and 120 billion parameters. Now, the
4. 0:09one that I really care about is the 20
5. 0:11billion parameter because that is the
6. 0:13one that I'm able to run it on my local
7. 0:16Mac machine. So, this is the one that I
8. 0:18really care about and I couldn't be more
9. 0:19excited about it because apparently it's
10. 0:21pretty good and that gives me full
11. 0:22privacy for my strategies. But of
12. 0:25course, what we really care about as
13. 0:26Algo traders is the performance of the
14. 0:28model for things such as writing
15. 0:30strategies in Python, debugging our
16. 0:33existing strategies, or if the model has
17. 0:35any actual trading knowledge. So, if all
18. 0:38that sounds interesting to you, let's
19. 0:39get right into the video. So, here's a
20. 0:41blog post released by OpenAI and it
21. 0:44gives us some information about the
22. 0:46models, but what I really care about is
23. 0:49the benchmark of it. So let's scroll
24. 0:51down and here we can see the GPD OSS 20
25. 0:57billion model without tools and here we
26. 1:00can see that it is beating slightly the
27. 1:0303 mini model. So that's pretty amazing
28. 1:05because I remember 03 mini used to be a
29. 1:08very good model. So if this is true that
30. 1:10is fantastic. Now to run this test we
31. 1:13have two ways. So one is to use JCGPT
32. 1:15which is our cloud version. It has all
33. 1:18the instructions for algo trading in
34. 1:20Python. And even if you don't have a GPU
35. 1:23of your own, you can just use this model
36. 1:25and you can check it out here. And this
37. 1:28new model, especially the 20 billion,
38. 1:30uses only one credit. So that is by far
39. 1:33the cheapest model on our list. So even
40. 1:36if you have a free account on Jess
41. 1:37trade, you can easily use this model.
42. 1:40But the whole point of doing this was
43. 1:42that we could have more privacy, right?
44. 1:44Because it's open source. So in this
45. 1:46video I'm not going to do that and
46. 1:48instead I will show you how to run a
47. 1:50completely offline version of it. So the
48. 1:52easiest way to do this is simply
49. 1:54download LM Studio which is great for
50. 1:57running local LLM models. Now I already
51. 2:00did this and in fact if you go to the
52. 2:02discover page you can see that GPOSS 20
53. 2:06billion is of this list and I already
54. 2:08downloaded this but because it's so new
55. 2:11only the GDUF version is available. So
56. 2:14the MLX version is now in case you don't
57. 2:16know what that is, the MLX version or
58. 2:19LLM models specifically for Apple
59. 2:21silicon machines. So if I run that
60. 2:23version which will be out approximately
61. 2:26tomorrow, I'm going to get 20 to 30%
62. 2:29better performance. So same machine but
63. 2:3120 or 30% better performance. So that's
64. 2:33really amazing. Now again I already
65. 2:35downloaded this. All right. So now we
66. 2:37are able to load the model. So let's go
67. 2:39here. Click here. Now, make sure to
68. 2:41increase the context length because by
69. 2:42default it is set to 4,000. That's very
70. 2:45low for our use case. So, I'm going to
71. 2:47set this to at least 30,000, but the max
72. 2:51supported value for this model is
73. 2:53131,000. So, that's pretty amazing. Now,
74. 2:56I'm seeing some new options on LM
75. 2:58Studio. I don't know if it's just for
76. 3:00this model or it came with the new
77. 3:02version of LM Studio, but to be honest
78. 3:04with you, I'm not sure how to change
79. 3:06these things. So, I'm just going to let
80. 3:07them what it is. It also gives us a
81. 3:11setting for number of experts. By
82. 3:12default, it is set to four, but I'm
83. 3:15really curious to know what would happen
84. 3:17if I increase this. So, anyways, let's
85. 3:19load the model. It's going to take a
86. 3:21while for this. All right. So, let's ask
87. 3:24it to
88. 3:27give me the entry and exit rules of a
89. 3:29golden cross strategy in Python.
90. 3:33Okay, so it's Python, not PyTorch. Let's
91. 3:36hit enter. And you see the model is
92. 3:38extremely fast and it is also a
93. 3:40reasoning model. So that's why it is
94. 3:42thinking before giving us the answer. So
95. 3:45it is giving me
96. 3:48the integrals.
97. 3:51It actually gave me some pseudo code. So
98. 3:54this is good. Then it is explaining the
99. 3:56code turning signals into position. So
100. 4:00yeah because you know by default it
101. 4:02doesn't know about Jesse framework. It
102. 4:04doesn't know how we're going to trade
103. 4:05this or back test it. So, this is
104. 4:07perfectly fine and pretty good.
105. 4:11Okay, so this seems really good. But
106. 4:12what I want to show you is this. You see
107. 4:14this value here? 66 token per second. So
108. 4:18that's the speed of the model. If I'm
109. 4:21able to run the MLX version, which will
110. 4:23come out tomorrow, I'm going to get like
111. 4:2580 or 90 token per second. So that's why
112. 4:29you would really care about the MLX
113. 4:31version. But anyway, so this is
114. 4:32perfectly acceptable. But now in order
115. 4:35to make sure the model is able to give
116. 4:38us answers for algo trading in
117. 4:40framework, which is something we can
118. 4:42actually back test or even live trade,
119. 4:45I'm going to have to go to this settings
120. 4:47page. And you see, you're going to have
121. 4:49to create a new preset and give it the
122. 4:51instructions. Now I'm going to paste the
123. 4:53whole instructions here. And it is
124. 4:55taking 20,000 tokens. So they are huge.
125. 4:59it. This has all the rules of the J
126. 5:02framework. It has a lot of examples in
127. 5:04it. So, this is really great. And in
128. 5:07case you're curious, yes, I'm going to
129. 5:09link to the repository that has this
130. 5:11instructions and leave it on the
131. 5:13description section of the video. Now,
132. 5:16the next thing is this model page. And
133. 5:20we need to change this temperature value
134. 5:22here. By default, I think this is set to
135. 5:240.5, but I have lowered it to 0.3. Now,
136. 5:29by the way, the temperature value is how
137. 5:31much you want the model to be creative.
138. 5:33So, the bigger it is, the more likely
139. 5:35the model is going to give you good
140. 5:37ideas, but it will also do more
141. 5:39hallucinations and I want to prevent
142. 5:41that. So, for that reason, I'm going to
143. 5:43set this to a value such as 0.3 or even
144. 5:460.2. And once you're done with this,
145. 5:48make sure to commit the changes. And now
146. 5:50if I go and ask the same thing from the
147. 5:53model. So let's go down here and simply
148. 5:56click on the regenerate message button.
149. 5:58You see now it is processing the prompt.
150. 6:01So it's taking a lot longer for it to
151. 6:03begin. And that is because we are now
152. 6:05feeding the model with 20,000 tokens of
153. 6:08input. So of course it's going to take
154. 6:10some time for the model to begin. But
155. 6:12once this initial prompt is over and it
156. 6:15starts giving us answers, my follow-up
157. 6:18questions of the model are going to
158. 6:20start almost instantly. So this is just
159. 6:22a one-time thing. Don't worry about it.
160. 6:24But still, even with that, you see when
161. 6:27it is printing the result, it is
162. 6:29significantly slower. Now it is slower,
163. 6:31but it is still fast enough. So I'm not
164. 6:33really sad about this. And what I'm
165. 6:37curious is the answer. So you see now it
166. 6:39is using the current syntax of the Jing
167. 6:41framework. And this is perfectly
168. 6:43correct. But I found the model to
169. 6:46sometimes write the strategy with a
170. 6:48weird syntax. So here it defined the
171. 6:50property, but it didn't go to the next
172. 6:52line and it just continued it like this.
173. 6:54I believe this should work, but it's a
174. 6:57bit odd. Now, another thing that I
175. 6:59noticed with this model is that it tends
176. 7:01to print a lot of comments and I really
177. 7:03like that. I think it's really helpful.
178. 7:05So let's wait for the whole thing. All
179. 7:08right. So now with the same prompt it
180. 7:10gave us an actual strategy that we are
181. 7:13able to run. So that's why these
182. 7:15instructions are super helpful. Now
183. 7:17let's go through the code very quickly.
184. 7:19So it defined two indicators. So one is
185. 7:22for moving average with the 20 period
186. 7:24and another with 50. The syntax is
187. 7:27simply trending EMA and then it is
188. 7:29passing the current candles which is a
189. 7:31perfectly correct syntax in Jesse. It
190. 7:33repeated that with the period of 50 in
191. 7:36the inter rule of the strategy which we
192. 7:38call shoot long in Jessie is it's saying
193. 7:41if it's the first index return false
194. 7:44because we don't have any previous data.
195. 7:46So this isn't correct actually. So this
196. 7:48shouldn't even be here but it is trying
197. 7:50to fetch the previous value of the EMA
198. 7:5220 with this syntax. This again is
199. 7:55incorrect because when you use this
200. 7:57syntax with Jesse, it will only give you
201. 7:59one value and that is the value of the
202. 8:02current indicator. So, it's not even a
203. 8:05list and this isn't correct. Now, this
204. 8:08is a bit weird because the previous
205. 8:09times that I ran this, it worked just
206. 8:11fine, but right now it's not. I'm not
207. 8:13sure why that is. Now, it is not
208. 8:14shorting. Okay, that's fine cuz we
209. 8:16didn't ask it. And for the update
210. 8:19position function, which only happens
211. 8:21whenever a new candle closes, it's
212. 8:23saying if it's a long position and the
213. 8:26EMA 20 is below EMA 50, which is kind of
214. 8:29a death cross. So in that case, liquid
215. 8:32the current position. This is correct
216. 8:34syntax. For the position sizing, it's
217. 8:36saying the entry is going to be the
218. 8:37current price. The stop is also going to
219. 8:40be the current entry minus two times of
220. 8:42the current ATR. This is correct. It is
221. 8:45risking 3% per each trade with the risk
222. 8:49to quantity utility function of Jesse.
223. 8:51This is perfectly correct. And to submit
224. 8:53the buy order, which is our entry order,
225. 8:54it's saying sell by equals quantity and
226. 8:57then the entry. Again, this is correct.
227. 8:59It's saying once the position is open,
228. 9:01if it's a long position, the stop loss
229. 9:02is going to be the quantity of the
230. 9:04current position. Again, correct syntax.
231. 9:06And the price of it is going to be the
232. 9:08current price minus two times of the
233. 9:11current ATR. So the two is here actually
234. 9:14and the take profit is going to be four
235. 9:16times of that. So it has a
236. 9:17risk-to-reward ratio of two. That's
237. 9:19really great. And then it defined the
238. 9:21watch list which is useful if you want
239. 9:23to trade the strategy live and see the
240. 9:25values in real time. So it's really
241. 9:28great. I like it. But now the token per
242. 9:31second is only 13. So it is
243. 9:33significantly lower but still acceptable
244. 9:36for a completely offline model. Now
245. 9:38guys, you know that I usually go through
246. 9:41all the questions and show you the
247. 9:43answers of the model. But in this video,
248. 9:46I'm not going to do that because I
249. 9:48actually did that. And after recording
250. 9:51the whole video, I realized that my
251. 9:53audio wasn't on. And right now,
252. 9:55unfortunately, I don't have time to
253. 9:57re-record the whole thing. So, I'm just
254. 9:59going to show you the ending result. And
255. 10:02I need you to trust me that I did run
256. 10:04these tests and these are the final
257. 10:06results that I was getting. So it got
258. 10:08most of the questions right. In fact, it
259. 10:10is getting perfect 10 in most of them.
260. 10:12Now with this one, it wasn't able to
261. 10:15complete the task which was to convert a
262. 10:16pin script strategy into Python code.
263. 10:19But whenever I ask it to write a
264. 10:20strategies, for example, write golden
265. 10:23cross strategy, it nailed it. I asked it
266. 10:25to develop a mean version of strategy
267. 10:27using Ballinger bands and the RSI with
268. 10:29this period of seven. It nailed that
269. 10:32question. I asked it to write a strategy
270. 10:34that uses Fibonacci retracement levels
271. 10:36for entry and exit points. It did it
272. 10:38beautifully and in all of these things
273. 10:40it was writing a lot of helpful comments
274. 10:43and that's something that I hadn't seen
275. 10:45in any other model. Then I asked it to
276. 10:48update my existing strategy code to add
277. 10:50some chart values. It wasn't able to do
278. 10:53a great job but it wasn't complete fail.
279. 10:55Then I asked it to write a trend
280. 10:57following strategy that uses 15 minutes
281. 10:59time frame for entry and the 4 time
282. 11:01frame for confirmation. So this is
283. 11:02something that we can only do with J
284. 11:04framework because if you try to use a
285. 11:07bigger time frame in let's say trading
286. 11:09views pin script you're going to hit the
287. 11:11look ahead bias which means you will be
288. 11:13using the price from the future and
289. 11:15you're going to get unrealistic results
290. 11:17but in Jesse that's not the case. Now
291. 11:19anyways the model was able to write a
292. 11:22very good strategy with that but I
293. 11:24needed to tweak one thing so get to get
294. 11:27it working and that's why I give it a 9
295. 11:29out of 10. Now for this question, I
296. 11:31asked it to improve my existing strategy
297. 11:34which was trading the 15 minutes time
298. 11:36frame. Now the strategy was losing money
299. 11:38because it was taking too many trades
300. 11:40when a lot of false breakouts were
301. 11:42happening. So I needed to know if the
302. 11:44model has some actual trading knowledge
303. 11:45to know what to do in these kind of
304. 11:47situations like what kind of filters it
305. 11:49will define to take less trades but only
306. 11:52trades that are actually worth it and it
307. 11:54wasn't helpful at all with that. So I
308. 11:56give it a zero. Next, I asked it to
309. 11:58develop a strategy that combines the
310. 12:00Williams percentage R, MACD, and
311. 12:02Ballinger bands. Now, I had to make two
312. 12:04changes to get it working, but it was
313. 12:06perfect. So, I give it eight. Next, I
314. 12:08asked it to write the entry functions of
315. 12:10the strategy for me in a way that it
316. 12:12should use two limit orders for opening
317. 12:14the position, and it should only risk
318. 12:162%age of capital per trade and it was
319. 12:19able to nail that. So it knows the
320. 12:21correct syntax for this framework to not
321. 12:23only use one order to open positions but
322. 12:26multiple. Next I asked it to write the
323. 12:28press reading strategy for ETH and BTC
324. 12:31and it was able to nail that question.
325. 12:34So overall these scores aren't as high
326. 12:36as the previous models that I reviewed
327. 12:38on this channel. But I believe this one
328. 12:40is still a very good one because all of
329. 12:43those models they had prompters bigger
330. 12:45than 100 billion at least. We're
331. 12:48actually able to run this one completely
332. 12:49locally and that is very valuable and
333. 12:52also with the questions that it didn't
334. 12:53get them working. I don't really care
335. 12:56because it still did a great job with
336. 12:58the other questions. So I believe this
337. 13:00model is perfectly usable on your local
338. 13:02machine if you are able to run it. And
339. 13:05even if you use a cloud version of it,
340. 13:07yes, you will lose some of your privacy,
341. 13:09but the pricing of it is significantly
342. 13:11low, which means that you can use it all
343. 13:14the time for research. So, I believe
344. 13:16OpenAI just did a wonderful job and I
345. 13:20wasn't expecting this at all. Like,
346. 13:22considering their previous releases and
347. 13:24how expensive those models were and not
348. 13:26actually really useful, I had no idea
349. 13:29that their open source model is going to
350. 13:31be this good. And by the way, recently I
351. 13:33also installed the coin tree coder
352. 13:36model. You know, the one that has lower
353. 13:38parameters and I was able to run it on
354. 13:40my MacBook machine and it was good, but
355. 13:43it wasn't nearly as good as this one. So
356. 13:45in case you are an open-source model
357. 13:48enthusiast like myself then you should
358. 13:50know this because again I think this is
359. 13:53the king of the open source models at
360. 13:55least the ones that you are able to run
361. 13:57them locally. Now there is another setup
362. 13:59that I'm really curious to try and that
363. 14:01is using an aentic framework for it. So
364. 14:04if we set up this model that way we
365. 14:07won't just ask the model to give us the
366. 14:09strategies copy and paste it into Jesse
367. 14:12and back test and see if it works or
368. 14:13not. it will be able to edit the source
369. 14:15of the strategy for us. So, it's super
370. 14:17helpful. So, if that's the kind of setup
371. 14:20that you want to see, let me know in the
372. 14:21comments and I might record a video just
373. 14:23about that. Now, if you guys want me to
374. 14:25make more videos like this, please like
375. 14:26the video, subscribe to the channel, and
376. 14:28give me your opinion in the comment
377. 14:30section. It helps me out a lot. Thank
378. 14:32you so much for watching, and I'll see
379. 14:33you in the next one.
380. 14:35[Music]
381. 14:38[Applause]
382. 14:40[Music]