Monte Carlo simulations for trading in Python is easy now — Transcript

Full transcript

1. 0:00Hey guys, it's Ole. It's been a while
2. 0:01since I recorded any videos because I've
3. 0:03been doing some serious development. I
4. 0:05wanted to tackle some of the most
5. 0:07important features that I've been
6. 0:08wanting to work for years. And today,
7. 0:11I'm happy to say that I implemented one
8. 0:13of the most important ones, which is
9. 0:15Monte Carlo simulations. They solve one
10. 0:17of the biggest questions in ago trading,
11. 0:19which is how to make sure your strategy
12. 0:21isn't overfit. So, let's get right into
13. 0:24it. Normally, when we do research, we
14. 0:27use back test, right? So let's say this
15. 0:29back testing result for one of my
16. 0:31strategies. It looks really well. But
17. 0:33before I go live with it, the serious
18. 0:35question is how do I ensure it's not
19. 0:37overfit? How do I ensure the risk that
20. 0:40I'm seeing here? For example, the max
21. 0:42throttle right now is minus almost 10%.
22. 0:45How do I make sure this is going to be
23. 0:47the result that I'm going to see in the
24. 0:49live trading environment? Because in
25. 0:51there it may be less. It might be minus
26. 0:545%. It may also be minus 20%. And these
27. 0:57three are actually very different. And
28. 0:59depending on those, I'm going to have to
29. 1:01adjust the position sizing of my
30. 1:02strategy because for instance, the win
31. 1:04rate in this back test result is 70%,
32. 1:07which is amazing. But what if that
33. 1:09wasn't the case? What if it was like
34. 1:1150%. Would my equity still look like
35. 1:14this or would it actually go down or
36. 1:16what would be the max roto number? So
37. 1:18these are some serious questions and to
38. 1:20answer them before this release, we had
39. 1:23no option. But starting today, we're
40. 1:25going to have Monte Carlo simulations
41. 1:27for that. But what is Monte Carlo
42. 1:29really? Well, to put it simply, Monte
43. 1:31Carlo basically means randomization. So,
44. 1:34what if we weren't actually so lucky in
45. 1:36the beginning of the back test like what
46. 1:38happened in here and we were actually so
47. 1:42unlucky like what happened in here. How
48. 1:44would our equity c look like then? Or
49. 1:47what if the price actually wasn't this?
50. 1:49What if there was a little bit noise in
51. 1:51it? So you see how market makers can
52. 1:53manipulate the order book. Well, how can
53. 1:56we ensure that our strategy is immune to
54. 1:58that? Well, one way is to do a little
55. 2:01bit of randomization. So for example,
56. 2:03what if we don't take the trade when
57. 2:05let's say the crossover happens now?
58. 2:07What if we take it when it happens a few
59. 2:10candles later? Now I know this sounds a
60. 2:12bit crazy, but it is actually a great
61. 2:14way for estimated stuff. And I emphasize
62. 2:16on the word estimation because when we
63. 2:18do Monte Carlo simulations, nothing is
64. 2:20going to be accurate anymore like the
65. 2:22actual back test itself. Everything is
66. 2:23going to be in estimations. But anyways,
67. 2:25don't let that scare you. It's actually
68. 2:26very simple. Now, if you look up Monte
69. 2:28Carlo simulations for trading on
70. 2:30YouTube, you're going to find a couple
71. 2:32of really good explanations of it and I
72. 2:34highly suggest you guys go and watch it.
73. 2:35Now, by the way, I am going to make my
74. 2:37own explanation video and a deep dive
75. 2:39into Monte Carlo in the future. But in
76. 2:42this video, I will just show you the
77. 2:43quickest way to run everything. So, in
78. 2:45the meanwhile, make sure to give these
79. 2:47videos a watch. But the point I want to
80. 2:49make is this that if you look it up,
81. 2:51you're going to see that they are using
82. 2:53Monte Carlo in three ways. Some people
83. 2:55like this guy are actually using it to
84. 2:57validate their trading signal. Now, if
85. 3:00you don't know what a trading signal or
86. 3:01an entry signal is, don't worry about it
87. 3:04because we're not going to work with it
88. 3:05now. But the other two types of videos
89. 3:07that I found are either shuffling the
90. 3:10trades or the actual candles. And if you
91. 3:12check out the documentation that I wrote
92. 3:14for Monte Carlo, and you can find it
93. 3:16under the research module and Monte
94. 3:18Carlo page, you can see I started by
95. 3:20explaining the difference between trade
96. 3:22order shuffling Monte Carlo and
97. 3:24candlesbased Monte Carlo. So, make sure
98. 3:26to give this page a serious read because
99. 3:28I worked really hard on it and it will
100. 3:30make things very clear for you. I also
101. 3:32included a lot of examples including
102. 3:34this one which will give you a complete
103. 3:37script to run just everything. Now if
104. 3:39you want to do this you need to ensure
105. 3:41the script is inside your Jessi project.
106. 3:43Now my project is named but so the file
107. 3:46that you run the script in it needs to
108. 3:48be in the root of your Jessie project.
109. 3:50Now assuming that's the case you only
110. 3:52care about one section and that is
111. 3:53configuration. We have trading routes.
112. 3:56So in this case it is this for me and
113. 3:59these are the data routes. So whatever
114. 4:01you put in your back testing page on
115. 4:03JS's dashboard make sure to put it
116. 4:05inside this file. And then we have the
117. 4:07simulation config. The number of
118. 4:08scenarios is really important. So I set
119. 4:11it to 200. You can set it to 100 or
120. 4:15maybe even 50 sometimes works, but the
121. 4:17higher it is, the more accurate numbers
122. 4:18you're going to get. So even if you can
123. 4:20set it to something like 1,000, it's
124. 4:22going to be better. But of course, it's
125. 4:24going to take longer for the simulation
126. 4:25to finish. Then we have the starting and
127. 4:27ending date of the simulation, the
128. 4:29progress bar, the benchmark, and the
129. 4:31fast mode. Now, I suggest you turn all
130. 4:33of these on. And then we have the
131. 4:35strategy config such as the starting
132. 4:36balance, the trading fees and this is an
133. 4:39important part which is the type of
134. 4:40Monte Carlo simulations you want to do
135. 4:42for candles. Now as I mentioned we have
136. 4:44two types of Monte Carlo in Jesse and
137. 4:46this is for the candle type. I'm going
138. 4:47to show you what that is in a second but
139. 4:49for now just so you know I suggest you
140. 4:51guys begin with moving block
141. 4:53bootstrapping method which is this one
142. 4:55and it simply accepts the batch size.
143. 4:57I've set it to one week like this. So
144. 5:00maybe do the same. But if you want to go
145. 5:02with the gajian noise option, which is
146. 5:04this one, you also need to pass other
147. 5:06values such as close sigma and other
148. 5:08sigma values. And if you don't know what
149. 5:10these are, again, I suggest you just
150. 5:12stick with this one because it doesn't
151. 5:14need as much configuration. It just
152. 5:16works out of the box. But it's going to
153. 5:17be different for every case. So if you
154. 5:19have the knowledge and the time, I
155. 5:20suggest playing around with both of
156. 5:22them. And that's it really. Now you just
157. 5:23need to go to your terminal and run the
158. 5:26command python. And the name of that
159. 5:28script file. I have named it test
160. 5:29Monteolo. UI and I already ran it and
161. 5:33here is the results. So let's begin with
162. 5:35the first type and that is the Monteol
163. 5:37candles and this is the one that I
164. 5:38actually care the most. Now with this
165. 5:40one what we actually do is that we add
166. 5:42some noise to the candles or in the case
167. 5:44of the moving block put strapping method
168. 5:46which I selected here. We're not adding
169. 5:48noise. We're just changing the order
170. 5:50which those movements happen in the
171. 5:52market. So for example, let's say the
172. 5:54market actually goes up 1% today and
173. 5:57tomorrow it goes down 2% and the next
174. 5:59day it goes up 3%. But what if the order
175. 6:02of it was different? What if it started
176. 6:04by going up 2% and then going down 2%
177. 6:06and then going up 1%. How would your
178. 6:09strategy behave if this was the case? So
179. 6:11you see I didn't add any noise and I
180. 6:13didn't make up price changes. I just
181. 6:16changed the order at which those price
182. 6:18changes happened in the market. Now, if
183. 6:21you're curious why this method actually
184. 6:22works and why it's a standard, just look
185. 6:24it up again. It is called moving block
186. 6:27bootstrapping. But that's what I went
187. 6:28with. And here's the thing. This is how
188. 6:30much my original practice made 81%. And
189. 6:34this is the max draw on. It was minus
190. 6:3510%. The sharp was 2.28.
191. 6:39Now, these numbers are really good, but
192. 6:41if you actually ran the Monte Carlo on
193. 6:43it and looked at the results here, this
194. 6:46is how it would have looked like. So you
195. 6:49see this green line here is the original
196. 6:51back test and these blue lines here are
197. 6:53the simulations. Now to put it very
198. 6:56simply, the lower the original back test
199. 6:59is among all of these simulations, the
200. 7:02better it is for you. It means that you
201. 7:04weren't as lucky in the original back
202. 7:08test and that's if you go live with the
203. 7:10same strategy, then the chances are that
204. 7:13you are actually going to get lucky in
205. 7:15the real trading. But if for example the
206. 7:18original back test was in here, I was
207. 7:21being super lucky in my back test and
208. 7:23there is a very low chance that in the
209. 7:26live trading I'm going to get the same
210. 7:28amount of luck. But because simulation
211. 7:31data isn't as accurate and as good as
212. 7:33the actual original data, I don't expect
213. 7:36this line to be really low, for example,
214. 7:38in here. Maybe it does happen for you at
215. 7:41some point and that may be a unicorn
216. 7:44strategy. I don't know. But I've never
217. 7:46found such a strategy. So how do we read
218. 7:48this? We have a table for it. So the
219. 7:51sharp ratio of the original back test
220. 7:53was 3.28.
221. 7:54But the median it was 2.33. Now what is
222. 7:58median? It is basically what was
223. 8:00happening in the middle of these
224. 8:01simulations. So you see the original
225. 8:03back test was better than the median and
226. 8:05that's not a great sign. But on the
227. 8:07other hand the best 5% the sharp was
228. 8:11almost five. So these are the best 5%
229. 8:14right? So yes, the strategy was a bit
230. 8:16lucky in the back test, but it wasn't as
231. 8:18lucky as the best 5%. Now there's also
232. 8:20another way to read this whole thing and
233. 8:22that is to say forget the original back
234. 8:25test. Let's just read the results from
235. 8:28the median and if that is good enough, I
236. 8:31will go live with it. So the sharp ratio
237. 8:33in the median is 2.33 and that is above
238. 8:37two which I personally consider really
239. 8:39good. So for that single reason I will
240. 8:42go live with this strategy. And I also
241. 8:44want to emphasize this. If you see some
242. 8:47good results in your Monte Carlo
243. 8:49simulations, yes, you can say there's a
244. 8:51good chance that your strategy isn't
245. 8:53overfit and that's great. But if you
246. 8:55don't see good results here, it doesn't
247. 8:57mean it is definitely overfit. So I also
248. 9:00want to emphasize on this. All right,
249. 9:02the Monte Carlo is not the law. It's not
250. 9:05telling you what is definitely
251. 9:07happening. It is just giving you some
252. 9:09estimations. But at the end of the day,
253. 9:11if you just had one single back test,
254. 9:13you didn't have any estimations
255. 9:15whatsoever. And having some estimations
256. 9:17is definitely better than not having
257. 9:20none at all. So from now on, when you do
258. 9:22some research, you develop a strategy.
259. 9:24If you like your back testing results
260. 9:26and now you're wondering, should I take
261. 9:28this live or not? Make sure to run a
262. 9:30Monte Carlo simulation on those results.
263. 9:33And that is going to tell you some
264. 9:35further answers which is really really
265. 9:37helpful. And yes, definitely this
266. 9:39feature will come to the UI dashboard of
267. 9:42Jessie which will make it super easy for
268. 9:44literally everybody to use it. But
269. 9:46that's going to take some more time. And
270. 9:48as I said in the beginning of the video,
271. 9:49I've been working on some serious
272. 9:51features. I cannot spoil anything just
273. 9:53yet, but just so you know, it is worth
274. 9:56the wait. So make sure to use the script
275. 9:58and give me some feedback. Thank you for
276. 10:00watching the video. I'll see you in the
277. 10:02next one.
278. 10:04[Music]
279. 10:07[Applause]
280. 10:07[Music]