I Gave Claude Code $200k to Trade Gold…(Full Build) — Transcript

Full transcript

1. 0:00So about a week ago I was doom scrolling
2. 0:02on Instagram and I've been seeing this
3. 0:04thing everywhere [music] that Claude is
4. 0:05incredibly good at investing. So I
5. 0:08wanted to put this to the test and I
6. 0:09started to thinking what if I could turn
7. 0:12Claude code into a gold coin.
8. 0:17>> [music]
9. 0:19>> And give this thing a full control over
10. 0:21my portfolio, have it running 24/7
11. 0:24constantly back testing strategies,
12. 0:26watching gold generating trades,
13. 0:28optimizing itself and tracking every
14. 0:30trade that I place so this thing learns
15. 0:32on its own. And for a moment it was
16. 0:34actually showing an 81% win rate which
17. 0:37is either absolutely insane or this
18. 0:39thing is completely broken. The idea was
19. 0:41simple if this thing could survive over
20. 0:43100 signals and actually show me
21. 0:45realistic figures that this thing's
22. 0:47going to work then I was going to give
23. 0:48it access to my 200k prop firm account.
24. 0:51Which may or may not have been a really
25. 0:53bad idea. But first I actually have to
26. 0:55build this thing which is way more
27. 0:57complicated than I originally thought
28. 0:59and took a lot more time. So when you're
29. 1:00trading gold you actually have to try
30. 1:02and figure out the direction of where
31. 1:03this thing is going by looking it at
32. 1:05different time frames. So all this means
33. 1:07is basically each one of these
34. 1:08candlesticks represents a certain period
35. 1:10of time. So right now each candlestick
36. 1:12represents 30 minutes. But for our
37. 1:14strategy specifically we wanted to look
38. 1:16at the 15 minute chart, the 1 hour
39. 1:18chart, the 4 hour chart and the daily
40. 1:20chart. And then this is going to allow
41. 1:22our system to have a full picture of
42. 1:24what's actually going on in the market
43. 1:26over different time frames. So the
44. 1:27system ended up looking a little bit
45. 1:29like this. So we have the gold chart, it
46. 1:31takes the last 200 candlesticks from the
47. 1:3315 minute chart, takes the last 200
48. 1:35candlesticks from the 1 hour chart, the
49. 1:364 hour chart and the daily. It then adds
50. 1:39this into a post gross database which is
51. 1:41basically just a database that we can
52. 1:42speak to. And I didn't just want to give
53. 1:44this thing one strategy because that
54. 1:46strategy might not work in all market
55. 1:48conditions. So instead I gave it four.
56. 1:50Strategy one hunts liquidity sweeps when
57. 1:52price wicks above a key level, grab stop
58. 1:55losses then reverses. Strategy two
59. 1:57follows trend pullbacks, classic EMA
60. 1:59continuation. And strategy three, wait
61. 2:01to volatility compression, then breakout
62. 2:03expansion. And strategy four, just sends
63. 2:05it in strong momentum. Each output has
64. 2:07an entry, a stop loss, a take profit, a
65. 2:09confidence score, and reasoning to why
66. 2:11it's took that trade. So, we've got the
67. 2:12data coming in, but now we need a way to
68. 2:14actually manage all this data. And I'm
69. 2:16going to call this section the brain.
70. 2:18So, we had our data coming in, we had
71. 2:20our strategies, but now we actually
72. 2:21needed to give this thing a real brain.
73. 2:23So, somewhere where we could actually
74. 2:25see all the results, all the testing
75. 2:26that it's doing. So, it built this. So,
76. 2:28every 30 minutes it runs a decision
77. 2:30pipeline. It expires stale signals,
78. 2:33ranks strategies by performance, like
79. 2:35you can see here. Liquidity sweeper's
80. 2:36coming up top right now. Checks your
81. 2:38risk limits, adjust your position size
82. 2:41based on volatility, blocks
83. 2:42overexposure, and limits your max daily
84. 2:45loss to 2%. And if everything passes,
85. 2:47and everything lines up between the
86. 2:48strategies, it will then send me a
87. 2:50signal over Telegram, and these are what
88. 2:52I'm going to be using to actually put
89. 2:53onto my prop firm account. No emotion,
90. 2:56no hesitation, just math. But, here's
91. 2:58where it actually gets interesting. So,
92. 3:00I introduced something called the Monte
93. 3:02Carlo simulation.
94. 3:03>> Monte Carlo simulation is a mathematical
95. 3:06technique which is used to estimate the
96. 3:07possible outcomes of an uncertain event.
97. 3:09So, take for example, the calculation of
98. 3:12the probability of rolling two standard
99. 3:15dice.
100. 3:17Well, if you wanted to calculate this
101. 3:19probability the brute force way, you
102. 3:21would have to roll the dice a whole
103. 3:23bunch, say 36,000
104. 3:26times if we consider that there are six
105. 3:28sides to a dice, we have two of them,
106. 3:30and we want to run this a thousand times
107. 3:32to get a good sample size.
108. 3:35But, with a Monte Carlo simulation, we
109. 3:37can reduce the number of rolls by
110. 3:39randomly sampling the possible outcomes,
111. 3:42knowing there are 36 combination of dice
112. 3:43rolls, and calculating the percentage of
113. 3:45times that we get, say, a seven.
114. 3:47>> So, every four hours it re-backtests
115. 3:50itself over a seven-day window, a 14-day
116. 3:52window, a 30-day window, and a 60-day
117. 3:54window. It then runs forward validation,
118. 3:5780% training and 20% testing. If our
119. 4:00sample performance collapses, then it
120. 4:01gets flagged as over-fitted. Then every
121. 4:036 hours it re-optimizes parameters, 80
122. 4:06combinations, Monte Carlo shuffle.
123. 4:08Basically, to see if randomness can beat
124. 4:10the original. And if it goes really
125. 4:12wrong, then it's going to throw the
126. 4:13parameters away. If it hits eight stop
127. 4:15losses in a row, the entire system shuts
128. 4:17down after 24 hours. No signals, no
129. 4:20trades, just silence. So, you can truly
130. 4:23sit with your pain. So, I decided to
131. 4:25give it a couple days to sort of do some
132. 4:27back test, come up with some data, and
133. 4:29then we could check on the metrics and
134. 4:31see how it was doing. But, when I went
135. 4:32to look at the metrics, [music]
136. 4:34something was a little bit off. The win
137. 4:36rate, 81%, sharp ratio, nine, over 1,500
138. 4:41trades in 30 days, which absolutely
139. 4:43makes no sense. Even the best quant
140. 4:45funds in the world can't produce that
141. 4:47consistency. So, either claw code is
142. 4:49crap gold or something was very wrong.
143. 4:52So, it's about four or five days later
144. 4:55now, and well, I've been placing all the
145. 4:58trades it's been telling me to place.
146. 5:00>> It's not looking good, babe.
147. 5:02It's not looking good.
148. 5:04>> So, as you can see, this is my live
149. 5:06funded account, and [music] we are
150. 5:09$11,500
151. 5:11down,
152. 5:12which is
153. 5:14terrible. We are cooked. Basically, we
154. 5:16just let this AI bot just absolutely
155. 5:18nuke our trading account. And it
156. 5:20actually took me to lose $11,000
157. 5:23just to figure out I was wrong with it.
158. 5:26So, none of this was actually making
159. 5:28sense. So, it's saying it had a 81% win
160. 5:30rate, you know, it's doing these back
161. 5:32tests, but somehow I'm $11,000 down. And
162. 5:35that's when I realized there was a huge
163. 5:37issue. So, it turns out that we were
164. 5:39accidentally fixing our own trades. So,
165. 5:42every back test the trade was being
166. 5:43simulated by the wrong bot. So, if a
167. 5:45signal was generated at $5,000 say, it
168. 5:48was tested at $5,100, which led to loads
169. 5:51of instant wins. So, I decided to build
170. 5:54this thing from scratch to make this
171. 5:56thing actually work. So, you can see a
172. 5:58little terrible diagram of it right now,
173. 6:01but the fast API handles the health
174. 6:03dashboard, the status, the candles, the
175. 6:05scheduler. So, this is where we got the
176. 6:08data from. So, this is all the
177. 6:09candlesticks. It refreshes every 15 to
178. 6:1160 minutes, 30 minutes, 90 seconds, and
179. 6:134 hours. Then it optimizes this data,
180. 6:16sends it through our pipeline, so
181. 6:18through the strategy detector. It then
182. 6:20picks which strategy to do, monitors the
183. 6:22outcome of the trades, does the trade
184. 6:25simulator, and does the metrics
185. 6:27calculator. And then all this data is
186. 6:28fed back into the Postgres database,
187. 6:31where things can then be optimized from
188. 6:32there. So, the external service, you've
189. 6:34got the 12 data API, got the API for
190. 6:36Telegram to send it to us. And then for
191. 6:38the dashboard, we've got the signal
192. 6:40history, the equity curve, the system
193. 6:41status, and the chart, which basically
194. 6:43shows us a lightweight version of what
195. 6:45we look on any trading chart website.
196. 6:47And don't worry if you don't understand
197. 6:49this. Basically, all you need to know is
198. 6:51one API for a website that can get us
199. 6:53all the candlestick data, price data,
200. 6:55the green and red things you see on a
201. 6:57chart. We get Claude code to build the
202. 6:59brain behind it. Does all these
203. 7:00calculations, you know, figures out our
204. 7:02strategies, does all that calculation
205. 7:03for us. You don't need to know a lot off
206. 7:05the top of your head. We're going to let
207. 7:06this thing do the work for it. We
208. 7:07connect it to Telegram, so we can
209. 7:08actually receive these signals over it.
210. 7:10And everything else, we let Claude
211. 7:11handle in the back end. You know, and
212. 7:13let me know in the comments if you
213. 7:14actually want me to make this GitHub uh
214. 7:16repo public, so you can all just access
215. 7:18it. If that's useful to you, let me know
216. 7:20in the comments. If you do actually want
217. 7:21to, you know, jump on a call me
218. 7:22one-to-one or speak to me one-to-one, I
219. 7:24do have a small community below where I
220. 7:25can speak to you directly. I can help
221. 7:27you with any of your builds. But, let's
222. 7:28get back to the video. So, where does
223. 7:30the system actually stand now? We've
224. 7:32actually got it showing realistic
225. 7:34measures. And right now, it's actually
226. 7:35showing a 60% win rate, which is
227. 7:38extremely good and a little bit
228. 7:40suspicious. But, I'm sure now that the
229. 7:42data is all working correctly. Is Claude
230. 7:44now a gold quant? Probably not. Is it
231. 7:47going to replace traders by the end of
232. 7:48the year? Probably yes. The metrics are
233. 7:50realistic. The system self-heals itself.
234. 7:53Does all the calculations for you. It
235. 7:55expects risk and can manage this and
236. 7:57shuts down during chaos. But the real
237. 8:00test is, now that this is fixed, to
238. 8:02actually see if we can recover this
239. 8:04$11,000 loss and actually make it back.
240. 8:07Because we need to actually give this
241. 8:08thing a chance and let it run. Now we
242. 8:10fix all the bugs. I'm going to let it
243. 8:12loose. And we're going to actually
244. 8:13either run this account into the ground
245. 8:15or make this money back and get the bag.
246. 8:17So, I'm going to again, I'm going to be
247. 8:19placing all the trades that it gives me
248. 8:21for the next week. And I'll catch up
249. 8:23with you for part two. Thanks for
250. 8:25watching, guys. And see you in the next
251. 8:26one.