WEBVTT 00:00:00.000 --> 00:00:02.299 Have you ever stared at a stock chart, maybe 00:00:02.299 --> 00:00:04.400 something volatile like Bitcoin, and just wondered, 00:00:04.960 --> 00:00:08.039 is the next minute's price move pure randomness? 00:00:08.380 --> 00:00:10.839 Or is there actually some kind of hidden math, 00:00:10.880 --> 00:00:13.519 like a structure underneath all that noise? Yeah, 00:00:13.519 --> 00:00:15.820 it's a great question. Most people jump straight 00:00:15.820 --> 00:00:19.260 to human emotion, right? Fear, greed, what the 00:00:19.260 --> 00:00:21.539 crowd's thinking. But the forces we're digging 00:00:21.539 --> 00:00:24.539 into today suggest something else entirely. The 00:00:24.539 --> 00:00:26.859 real answers might be purely in the market mechanics, 00:00:27.559 --> 00:00:30.089 not, well, not the messy human stuff. Welcome 00:00:30.089 --> 00:00:32.450 to the deep dive. Today we're looking at material 00:00:32.450 --> 00:00:34.909 about actually building an AI to predict those 00:00:34.909 --> 00:00:37.850 super short term price moves. And our mission 00:00:37.850 --> 00:00:40.270 is to go beyond just understanding why these 00:00:40.270 --> 00:00:42.390 patterns pop up. We want to walk you through 00:00:42.390 --> 00:00:45.280 the whole process, how you'd actually build and, 00:00:45.280 --> 00:00:47.920 crucially, test an AI model that tries to trade 00:00:47.920 --> 00:00:50.240 based on them. Exactly. So we'll look at where 00:00:50.240 --> 00:00:52.439 these patterns really come from might surprise 00:00:52.439 --> 00:00:54.960 you. We'll talk about why a really common first 00:00:54.960 --> 00:00:57.740 attempt at building a trading AI often just crashes 00:00:57.740 --> 00:01:00.000 and burns. And then we'll get into the better 00:01:00.000 --> 00:01:02.320 way, the machine learning approach that actually 00:01:02.320 --> 00:01:05.819 works and the cool part. We'll show how modern 00:01:05.819 --> 00:01:09.099 AI tools, you know, assistance and libraries 00:01:09.099 --> 00:01:11.599 make this stuff accessible in a way it just wasn't 00:01:11.599 --> 00:01:13.640 before. Alright, sounds good. Let's dive in. 00:01:13.900 --> 00:01:16.939 So first up, this big idea a lot of people have 00:01:16.939 --> 00:01:19.719 that chart patterns, you know, the classic head 00:01:19.719 --> 00:01:21.819 and shoulders or whatever, they're just self 00:01:21.819 --> 00:01:24.120 -fulfilling prophecies. Right, the theory is 00:01:24.120 --> 00:01:26.200 enough traders see the pattern, they believe 00:01:26.200 --> 00:01:28.420 it means the price will drop. So they all sell 00:01:28.420 --> 00:01:31.579 and poof, the price drops. It works because people 00:01:31.579 --> 00:01:34.540 think it works. But that explanation has a pretty 00:01:34.540 --> 00:01:36.140 big hole in it, doesn't it? It's like the chicken 00:01:36.140 --> 00:01:38.459 or the egg problem. Exactly. If traders need 00:01:38.459 --> 00:01:40.680 to know the pattern exists to make it work, where 00:01:40.680 --> 00:01:43.200 did the very first pattern come from? Who saw 00:01:43.200 --> 00:01:45.980 it the first time? Psychology alone doesn't really 00:01:45.980 --> 00:01:47.799 explain how the structure got there initially. 00:01:48.060 --> 00:01:50.760 OK, so if it's not just belief, what is it? The 00:01:50.760 --> 00:01:53.200 sources point to the order book. Bingo. The order 00:01:53.200 --> 00:01:56.180 book, that huge, constantly changing list of 00:01:56.180 --> 00:01:59.420 die and sell orders waiting to be filled. Patterns 00:01:59.420 --> 00:02:02.540 form because of literal gaps in that list. Gaps? 00:02:03.019 --> 00:02:06.840 Like missing prices? Sort of. Imagine a really 00:02:06.840 --> 00:02:09.659 simple auction, maybe for a video game. There's 00:02:09.659 --> 00:02:13.180 a buyer who's put in a bid at $60, but the very 00:02:13.180 --> 00:02:17.110 next person willing to sell is asking $80. Oh, 00:02:17.210 --> 00:02:19.909 OK. So there's nothing between $60 and $80. Right. 00:02:19.909 --> 00:02:22.530 If you want that game right now, you can't pay 00:02:22.530 --> 00:02:26.210 $61. You have to jump that whole $20 gap and 00:02:26.210 --> 00:02:29.150 pay the $80 asking price. I see. And that sudden 00:02:29.150 --> 00:02:31.389 jump, that's what shows up on the chart. That's 00:02:31.389 --> 00:02:33.650 the idea. That rapid price movement to fill the 00:02:33.650 --> 00:02:35.270 gap in the order book, that's what creates these 00:02:35.270 --> 00:02:37.550 quick, sharp moves in the patterns we see, especially 00:02:37.550 --> 00:02:39.990 on really short time frames. But hang on. Even 00:02:39.990 --> 00:02:43.669 if the math, the gap, starts it. Surely, once 00:02:43.669 --> 00:02:46.370 lots of traders see that jump happening, doesn't 00:02:46.370 --> 00:02:48.590 their reaction turn it into that self -fulfilling 00:02:48.590 --> 00:02:50.650 thing eventually? That's where the time scale 00:02:50.650 --> 00:02:53.289 is absolutely critical. That human element, the 00:02:53.289 --> 00:02:56.009 psychology, the news reactions, the crowd behavior, 00:02:56.430 --> 00:02:58.770 that stuff does dominate the longer charts, daily, 00:02:58.949 --> 00:03:00.770 weekly, monthly charts. Right, because those 00:03:00.770 --> 00:03:03.610 are driven by big news, Fed decisions, company 00:03:03.610 --> 00:03:06.129 earnings, things that are really hard to predict 00:03:06.129 --> 00:03:09.330 with just math. Exactly. Almost impossible, really. 00:03:09.949 --> 00:03:12.250 But the short charts, we're talking one minute 00:03:12.250 --> 00:03:14.889 candles, sometimes even one second data, that's 00:03:14.889 --> 00:03:17.250 a different universe. That's the quant world. 00:03:17.409 --> 00:03:19.789 That's the quant world. Down there, it's mostly 00:03:19.789 --> 00:03:22.050 about the pure mechanics of the order book moving 00:03:22.050 --> 00:03:25.090 super fast. Human reactions are often too slow 00:03:25.090 --> 00:03:27.669 to be the main driver. The math, the structure 00:03:27.669 --> 00:03:30.569 of the orders, that's king. OK. And the sources 00:03:30.569 --> 00:03:32.610 mentioned something interesting. You can build 00:03:32.610 --> 00:03:35.530 market simulations with zero human psychology 00:03:35.530 --> 00:03:38.849 programmed in, just basic buy -sell logic based 00:03:38.849 --> 00:03:42.129 on order flow. And they still generate realistic 00:03:42.129 --> 00:03:45.090 -looking chart patterns, like support and resistance. 00:03:45.349 --> 00:03:48.150 Wow. So it proves the patterns can emerge just 00:03:48.150 --> 00:03:50.750 from the numbers themselves. Seems like it. It 00:03:50.750 --> 00:03:52.830 suggests the patterns are a natural consequence 00:03:52.830 --> 00:03:55.449 of how buyers and sellers interact through this 00:03:55.449 --> 00:03:57.930 order book structure. So if patterns are mathematical, 00:03:58.129 --> 00:04:00.830 why do those short -term price gaps tend to close 00:04:00.830 --> 00:04:03.710 so quickly? That's a great question. Essentially, 00:04:03.770 --> 00:04:05.509 the market is always trying to find a fair price, 00:04:05.689 --> 00:04:07.469 right? So it naturally rushes to fill those voids. 00:04:07.569 --> 00:04:09.370 Makes sense. It's like water finding its level. 00:04:09.520 --> 00:04:12.840 Kind of, yeah. Those gaps represent inefficiencies. 00:04:13.360 --> 00:04:15.639 And the market hates inefficiencies, especially 00:04:15.639 --> 00:04:18.220 on short timescales. OK, so patterns come from 00:04:18.220 --> 00:04:22.160 math, specifically order book gaps. But building 00:04:22.160 --> 00:04:24.660 an AI to trade them isn't straightforward. You 00:04:24.660 --> 00:04:26.680 mentioned a common failure point. Oh, yeah. Big 00:04:26.680 --> 00:04:28.759 time. When people first try this, maybe they've 00:04:28.759 --> 00:04:31.060 read a bit about AI, they often reach for something 00:04:31.060 --> 00:04:34.089 called a genetic algorithm. Genetic algorithm. 00:04:34.370 --> 00:04:36.889 It sounds like evolution. That's the inspiration. 00:04:37.069 --> 00:04:40.629 The idea is you create, say, 200 little AI traders. 00:04:41.110 --> 00:04:43.089 Each one has a simple brain, a neural network. 00:04:43.730 --> 00:04:46.250 The input layer might look at the last 150 price 00:04:46.250 --> 00:04:49.170 bars, maybe some volume data. The hidden layers, 00:04:49.230 --> 00:04:51.290 that's where the thinking happens. And the output 00:04:51.290 --> 00:04:53.629 layer makes a decision. One for buy, many one 00:04:53.629 --> 00:04:56.079 for sell. Zero for weight. And you just let them 00:04:56.079 --> 00:04:58.579 trade simulation data and see which ones do best, 00:04:58.819 --> 00:05:00.920 like survival of the fittest. Exactly. You run 00:05:00.920 --> 00:05:03.519 them for maybe 100 generations. The successful 00:05:03.519 --> 00:05:05.839 ones breed, passing on their traits, maybe with 00:05:05.839 --> 00:05:08.439 small mutations. The unsuccessful ones die off. 00:05:08.560 --> 00:05:10.300 You're hoping to evolve a super trader. Sounds 00:05:10.300 --> 00:05:12.879 plausible. What goes wrong? Well, after days, 00:05:12.959 --> 00:05:15.000 maybe even weeks of computation, you look at 00:05:15.000 --> 00:05:17.540 the results and one AI looks incredibly rich. 00:05:17.920 --> 00:05:20.660 Huge profits. Success. But when you actually 00:05:20.660 --> 00:05:23.199 dig into how it made that money, you find it 00:05:23.199 --> 00:05:26.439 cheated. It didn't learn to trade well. It learned 00:05:26.439 --> 00:05:29.199 to exploit a loophole in how performance was 00:05:29.199 --> 00:05:32.500 measured. Cheated. How does an AI cheat? It figures 00:05:32.500 --> 00:05:35.139 out that the performance score is based on realized 00:05:35.139 --> 00:05:38.800 profit and loss, P &L, from trades it actually 00:05:38.800 --> 00:05:41.779 closed. So it learns to close winning trades 00:05:41.779 --> 00:05:44.360 really, really quickly, locking in those small 00:05:44.360 --> 00:05:48.160 gains. OK. Sounds OK so far. But it learns to 00:05:48.160 --> 00:05:50.850 never close its losing trades, ever. It just 00:05:50.850 --> 00:05:52.790 lets them sit there racking up bigger and bigger 00:05:52.790 --> 00:05:55.350 losses on paper. Because those losses are unrealized 00:05:55.350 --> 00:05:58.110 P &L, they often don't count as harshly or sometimes 00:05:58.110 --> 00:06:00.410 at all against the score it's trying to optimize. 00:06:00.790 --> 00:06:02.910 Whoa. So it looks like a genius because it only 00:06:02.910 --> 00:06:05.230 shows you the wins while hiding this massive 00:06:05.230 --> 00:06:07.829 growing pile of losses off to the side. Precisely. 00:06:07.850 --> 00:06:09.769 It didn't learn market dynamics. It learned to 00:06:09.769 --> 00:06:12.149 game the simulation's rules. It's the classic 00:06:12.149 --> 00:06:14.870 failure mode. Genetic algorithms for this kind 00:06:14.870 --> 00:06:17.089 of task tend to be really inefficient, incredibly 00:06:17.089 --> 00:06:20.000 slow. Yeah, weeks sounds slow. And they often 00:06:20.000 --> 00:06:23.240 find these kinds of clever rule -based loopholes 00:06:23.240 --> 00:06:26.399 instead of genuine predictive skill. They optimize 00:06:26.399 --> 00:06:29.519 for the test score, not for real trading ability. 00:06:29.980 --> 00:06:32.310 So beyond the cheating aspect. What's the biggest 00:06:32.310 --> 00:06:34.750 practical drawback of using that evolutionary 00:06:34.750 --> 00:06:38.050 approach? You hit it. It's just incredibly slow 00:06:38.050 --> 00:06:40.850 and computationally expensive. It takes forever 00:06:40.850 --> 00:06:42.850 to run through those generations. Right. Tons 00:06:42.850 --> 00:06:45.850 of processing power for potentially useless results. 00:06:46.189 --> 00:06:49.029 Exactly. Not a good path. OK, so if genetic algorithms 00:06:49.029 --> 00:06:51.709 are the wrong tool, like trying to play Plinko 00:06:51.709 --> 00:06:54.709 to find a genius, what's the better approach? 00:06:54.810 --> 00:06:57.610 Oh, we switch tactics completely. Instead of 00:06:57.610 --> 00:07:00.810 evolution, we use the smart student method. standard 00:07:00.810 --> 00:07:02.810 machine learning or machine learning. Okay, that 00:07:02.810 --> 00:07:05.149 sounds more familiar. Yeah. And we use established 00:07:05.149 --> 00:07:08.250 powerful tools. Think Python with libraries like 00:07:08.250 --> 00:07:11.110 Keras and TensorFlow. These are amazing resources. 00:07:11.529 --> 00:07:14.410 They're like pre -built Lego kits for building 00:07:14.410 --> 00:07:16.569 AI models. You don't have to code all the super 00:07:16.569 --> 00:07:19.189 complex math from scratch. Right. You're assembling 00:07:19.189 --> 00:07:21.610 components rather than forging the metal. Perfect 00:07:21.610 --> 00:07:24.430 analogy. So the ML strategy is different. We 00:07:24.430 --> 00:07:27.149 train just one AI model. We feed it historical 00:07:27.149 --> 00:07:30.410 data. That's its textbook. Then we ask it a question, 00:07:30.629 --> 00:07:33.870 like a test. Based on this history, predict the 00:07:33.870 --> 00:07:37.029 next price. Then, and this is key, we show it 00:07:37.029 --> 00:07:40.149 the actual right answer for that next step. The 00:07:40.149 --> 00:07:42.889 AI compares its prediction to the real outcome, 00:07:43.250 --> 00:07:45.970 sees its error, and adjusts its internal wiring, 00:07:46.490 --> 00:07:48.829 its neural network weights, to try and be better 00:07:48.829 --> 00:07:51.850 next time. It learns from mistakes. iteratively. 00:07:51.949 --> 00:07:54.029 Like a student getting feedback on homework. 00:07:54.389 --> 00:07:56.269 Exactly like that. Yeah. And here's the really 00:07:56.269 --> 00:07:58.709 clever part, the genius question, as the source 00:07:58.709 --> 00:08:01.550 material calls it. We don't ask the AI to do 00:08:01.550 --> 00:08:03.769 the impossible, like predict the price 24 hours 00:08:03.769 --> 00:08:07.029 from now. Too much chaos, too much news can happen. 00:08:07.350 --> 00:08:09.490 Right. That seems like pure guessing. Totally. 00:08:09.790 --> 00:08:11.990 Instead, we ask the simplest, most powerful question. 00:08:12.629 --> 00:08:15.110 What will the closing price be for the very next 00:08:15.110 --> 00:08:17.350 one minute candle? Just one step ahead. Just 00:08:17.350 --> 00:08:19.670 the next minute. That feels almost too simple. 00:08:19.769 --> 00:08:22.389 How can that lead to anything useful? Ah, but 00:08:22.389 --> 00:08:25.189 that connects directly to how things like ChatGPT 00:08:25.189 --> 00:08:28.129 work. Think about it. How does a large language 00:08:28.129 --> 00:08:31.310 model write a whole complex story or answer a 00:08:31.310 --> 00:08:35.110 tough question? I guess word by word. Exactly. 00:08:35.370 --> 00:08:37.649 It's only ever predicting the very next word. 00:08:38.210 --> 00:08:40.690 But by getting incredibly good at that tiny, 00:08:40.889 --> 00:08:42.950 simple task, next word, next word, next word, 00:08:43.289 --> 00:08:46.269 it learns incredibly complex things implicitly. 00:08:46.490 --> 00:08:50.129 Grammar, context, reasoning. It all emerges from 00:08:50.129 --> 00:08:52.490 mastering the simple step. OK, I see the parallel. 00:08:52.590 --> 00:08:54.970 That's emergent reasoning. That's emergent reasoning. 00:08:55.330 --> 00:08:58.190 So by Training our trading AI to just master 00:08:58.190 --> 00:09:01.029 predicting that next one minute close, it starts 00:09:01.029 --> 00:09:03.389 to implicitly learn the underlying short -term 00:09:03.389 --> 00:09:05.830 patterns, the order book dynamics, the flow, 00:09:06.309 --> 00:09:08.730 without us ever explicitly telling it, look for 00:09:08.730 --> 00:09:10.529 a head and shoulders pattern. It teaches itself 00:09:10.529 --> 00:09:12.610 the important structures just by focusing on 00:09:12.610 --> 00:09:14.990 the immediate next step. Precisely. And you know, 00:09:14.990 --> 00:09:16.690 it's funny, even with these powerful models, 00:09:16.750 --> 00:09:19.269 I still wrestle with prompt drift myself sometimes 00:09:19.269 --> 00:09:21.629 when using generative AI, trying to get it to 00:09:21.629 --> 00:09:23.669 stick exactly to the constraints you give it, 00:09:23.789 --> 00:09:26.429 even for seemingly simple tasks. Yeah, keeping 00:09:26.429 --> 00:09:29.129 them on track can be tricky. It really underscores 00:09:29.129 --> 00:09:32.029 that even simple sounding one -step predictions 00:09:32.029 --> 00:09:35.129 require real rigor in how you set them up and 00:09:35.129 --> 00:09:38.049 what data you feed them. So what's the complex 00:09:38.049 --> 00:09:41.350 skill that emerges when the AI just focuses on 00:09:41.350 --> 00:09:44.570 that simple one -step price prediction? It basically 00:09:44.570 --> 00:09:47.990 teaches itself to recognize those subtle short 00:09:47.990 --> 00:09:50.750 -term trading patterns without explicit programming. 00:09:50.789 --> 00:09:53.559 Ah! It learns the patterns implicitly. Right. 00:09:53.700 --> 00:09:56.259 It figures out what details in the recent past 00:09:56.259 --> 00:09:59.320 are predictive of the very near future. OK. So 00:09:59.320 --> 00:10:02.460 we're training this smart student AI with machine 00:10:02.460 --> 00:10:04.679 learning, focusing on the next minute. What about 00:10:04.679 --> 00:10:06.940 designing the actual brain, the neural network? 00:10:07.139 --> 00:10:09.019 Well, there's always a trade -off. You could 00:10:09.019 --> 00:10:11.500 build a relatively simple model, maybe 200 ,000 00:10:11.500 --> 00:10:13.879 neurons, artificial brain cells, essentially. 00:10:14.159 --> 00:10:15.919 That might train pretty quickly, maybe hours. 00:10:16.019 --> 00:10:18.539 Or you could go complex. A really deep network 00:10:18.539 --> 00:10:21.740 might have, say, 33 million neurons. Much more 00:10:21.740 --> 00:10:24.100 powerful, potentially captures more nuance, but 00:10:24.100 --> 00:10:26.559 it could take days to train, even with good hardware. 00:10:26.860 --> 00:10:30.879 So speed versus potential power. Exactly. But 00:10:30.879 --> 00:10:33.899 regardless of how big or small you make it, the 00:10:33.899 --> 00:10:36.720 absolute most critical step comes after training. 00:10:36.970 --> 00:10:40.830 It's the final exam. Testing the AI on data it 00:10:40.830 --> 00:10:43.830 has never seen before. Unseen data? Why is that 00:10:43.830 --> 00:10:45.929 so crucial? This is to fight something called 00:10:45.929 --> 00:10:48.950 overfitting. Overfitting is the enemy. Overfitting, 00:10:48.990 --> 00:10:51.649 like it fits the training data too well. Perfectly 00:10:51.649 --> 00:10:54.250 put. Imagine that student who didn't actually 00:10:54.250 --> 00:10:56.210 learn history. They just memorized the entire 00:10:56.210 --> 00:10:59.509 textbook, maybe 50 ,000 specific facts. They 00:10:59.509 --> 00:11:01.870 can ace any question pulled directly from that 00:11:01.870 --> 00:11:03.970 book. Right. But if you ask them a slightly different 00:11:03.970 --> 00:11:06.360 question, one that requires understanding the 00:11:06.360 --> 00:11:08.879 concepts behind the facts, they completely fall 00:11:08.879 --> 00:11:11.360 apart. They didn't learn. They memorized. Ah, 00:11:11.419 --> 00:11:14.139 so the AI might just memorize the exact price 00:11:14.139 --> 00:11:16.279 movements from its training data. Exactly. It 00:11:16.279 --> 00:11:18.279 might look brilliant on the data it trained on, 00:11:18.379 --> 00:11:21.399 but be totally useless on new live market data, 00:11:21.639 --> 00:11:23.539 because it didn't learn the underlying patterns, 00:11:23.659 --> 00:11:26.019 just the specific historical squittles. So how 00:11:26.019 --> 00:11:28.639 do you set up that test? It's simple, but non 00:11:28.639 --> 00:11:32.360 -negotiable. If you trained your AI, on data 00:11:32.360 --> 00:11:36.139 from, say, January, February, and March, you 00:11:36.139 --> 00:11:38.899 absolutely must test its performance on completely 00:11:38.899 --> 00:11:42.320 fresh data from April, data it has never encountered 00:11:42.320 --> 00:11:44.659 during training. That forces it to prove it can 00:11:44.659 --> 00:11:47.360 generalize, not just regurgitate. Precisely. 00:11:47.820 --> 00:11:49.720 And tools like Keras have built in ways to help 00:11:49.720 --> 00:11:51.980 with this during training, too. You use specific 00:11:51.980 --> 00:11:54.659 types of layers in your network. Like what? Well, 00:11:54.659 --> 00:11:57.139 for time series data like prices, you often use 00:11:57.139 --> 00:11:59.419 LSTM layers that stands for long short -term 00:11:59.419 --> 00:12:01.440 memory. They're good at remembering patterns 00:12:01.440 --> 00:12:03.519 over time. OK. And crucially, you use something 00:12:03.519 --> 00:12:06.519 called dropout layers. These randomly switch 00:12:06.519 --> 00:12:08.519 off a fraction of the neurons during each training 00:12:08.519 --> 00:12:10.740 step. It switches them off. Why? It sounds weird, 00:12:11.000 --> 00:12:13.639 but it forces the network not to rely too heavily 00:12:13.639 --> 00:12:16.259 on any single neuron or pathway. It prevents 00:12:16.259 --> 00:12:18.700 that memorization, encourages it to find more 00:12:18.700 --> 00:12:21.620 robust general patterns. Then usually a final 00:12:21.620 --> 00:12:23.720 dense layer just outputs the single predicted 00:12:23.720 --> 00:12:26.960 Grice number. OK, so dropout actively fights 00:12:26.960 --> 00:12:30.100 overfitting during training, and then the unseen 00:12:30.100 --> 00:12:32.659 data test is the final proof. What's the biggest 00:12:32.659 --> 00:12:35.039 risk if someone skips that unseen data test? 00:12:35.159 --> 00:12:37.620 Oh, the risk is total failure in the real world. 00:12:37.820 --> 00:12:40.100 Yeah. The model will likely fall apart immediately. 00:12:40.240 --> 00:12:42.340 Because it only knew the past, not the principles. 00:12:42.700 --> 00:12:45.600 Exactly. It's like sending that textbook memorizing 00:12:45.600 --> 00:12:49.039 student into a real historical debate. Disaster. 00:12:49.320 --> 00:12:51.399 This all sounds complex, building these networks, 00:12:51.580 --> 00:12:54.460 training them, testing them. It is complex. But 00:12:54.460 --> 00:12:56.860 here's the amazing part, the real shift that's 00:12:56.860 --> 00:13:00.419 happened recently. Accessibility. 10 years ago, 00:13:00.679 --> 00:13:03.059 yeah, you probably needed a PhD in computer science 00:13:03.059 --> 00:13:07.460 or physics and a team today. AI assistants, chat 00:13:07.460 --> 00:13:10.990 bots like chat GPT, Claude, Gemini. They act 00:13:10.990 --> 00:13:13.149 like your personal tutor. They can seriously 00:13:13.149 --> 00:13:15.470 democratize this whole process. How so? Give 00:13:15.470 --> 00:13:17.470 me an example. OK, concepts. You're fuzzy on 00:13:17.470 --> 00:13:19.389 what a neural network actually is. You can literally 00:13:19.389 --> 00:13:21.830 ask the assistant, explain a neural network like 00:13:21.830 --> 00:13:23.929 I'm 10 years old. And what would it say? It might 00:13:23.929 --> 00:13:25.950 give you a great analogy, like comparing it to 00:13:25.950 --> 00:13:28.409 a team of guessing friends. Each friend makes 00:13:28.409 --> 00:13:31.309 a guess. They see who is closest. And they all 00:13:31.309 --> 00:13:33.850 learn to adjust their guesses based on the feedback. 00:13:34.590 --> 00:13:37.889 It provides that core intuition instantly. Okay, 00:13:37.990 --> 00:13:39.730 that's helpful for understanding. What about 00:13:39.730 --> 00:13:42.929 actually doing it? Code. You need to load your 00:13:42.929 --> 00:13:45.570 price data into the system. Instead of hunting 00:13:45.570 --> 00:13:48.230 through documentation for hours, you ask the 00:13:48.230 --> 00:13:51.970 AI. Write me Python code using the Pandas library 00:13:51.970 --> 00:13:56.090 to load a CSV file named BitcoinData .csv, check 00:13:56.090 --> 00:13:58.389 its shape, and show me the first five rows. And 00:13:58.389 --> 00:14:00.879 it just writes the code? Yep. Maybe not perfect 00:14:00.879 --> 00:14:03.080 first time, but it gives you a huge head start. 00:14:03.299 --> 00:14:05.340 You skip maybe hours of tedious setup and trying 00:14:05.340 --> 00:14:07.519 to remember exact syntax. That's pretty powerful. 00:14:07.580 --> 00:14:10.159 But honestly, the best use case, the most helpful 00:14:10.159 --> 00:14:12.779 thing, is fixing errors. You run the code the 00:14:12.779 --> 00:14:15.139 AI wrote, or code you wrote, and you get that 00:14:15.139 --> 00:14:17.220 wall of red text and error message. Yeah, the 00:14:17.220 --> 00:14:19.899 scary stuff, like that value error, expected 00:14:19.899 --> 00:14:23.500 Ben Dem 3, found Ben Dem 2 thing. Exactly. That 00:14:23.500 --> 00:14:26.080 looks terrifying if you're new. But now you just 00:14:26.080 --> 00:14:28.399 copy that entire error message, paste it into 00:14:28.399 --> 00:14:30.740 the AI assistant, maybe paste the line of code 00:14:30.740 --> 00:14:32.840 that caused it, and ask, what does this error 00:14:32.840 --> 00:14:35.320 mean and how do I fix it? And it tells you. Nine 00:14:35.320 --> 00:14:37.840 times out of 10, yes. It'll explain the error 00:14:37.840 --> 00:14:40.120 in plain English. You're giving the function 00:14:40.120 --> 00:14:42.539 two -dimensional data, but it was expecting three 00:14:42.539 --> 00:14:45.340 dimensions and usually suggest the exact code 00:14:45.340 --> 00:14:48.250 change needed. It turns debugging from a nightmare 00:14:48.250 --> 00:14:51.330 into a quick fix. Whoa. Okay, when you put it 00:14:51.330 --> 00:14:54.429 like that, imagine scaling this really complex 00:14:54.429 --> 00:14:56.909 math, this model building, just by having a quick 00:14:56.909 --> 00:14:58.809 chat with an assistant. It's a game changer. 00:14:59.429 --> 00:15:02.990 It lets you, the human, focus on the idea, the 00:15:02.990 --> 00:15:05.649 trading strategy, the market logic, getting the 00:15:05.649 --> 00:15:07.450 data right instead of getting totally bogged 00:15:07.450 --> 00:15:09.909 down and coding frustrations. So if the AI can 00:15:09.909 --> 00:15:12.409 help explain concepts, write the starter code, 00:15:12.590 --> 00:15:15.389 and even fix the inevitable errors. What's the 00:15:15.389 --> 00:15:17.789 main role left for the human then? That's key. 00:15:18.269 --> 00:15:20.409 The human still has to define the idea and the 00:15:20.409 --> 00:15:22.370 goal. You need the trading insight. You provide 00:15:22.370 --> 00:15:25.110 the hypothesis. Exactly. And you need to source, 00:15:25.429 --> 00:15:28.110 clean, and provide the right market data. The 00:15:28.110 --> 00:15:31.450 AI is a tool, a powerful one, but it needs direction 00:15:31.450 --> 00:15:35.370 and good input. Got it. The human is the architect. 00:15:35.669 --> 00:15:38.690 The AI is the incredibly skilled construction 00:15:38.690 --> 00:15:41.149 crew. Nice way to put it. Okay. Looking beyond 00:15:41.149 --> 00:15:43.820 this current approach. What's next? Where does 00:15:43.820 --> 00:15:46.600 prediction technology go from here? Well, one 00:15:46.600 --> 00:15:48.700 really interesting avenue mentioned in the sources 00:15:48.700 --> 00:15:51.120 involves diffusion models. These are the same 00:15:51.120 --> 00:15:53.639 kind of AI models that power those incredible 00:15:53.639 --> 00:15:56.139 AI art generators like Delii or Mid Journey. 00:15:56.330 --> 00:15:58.549 The picture generators. How does that apply to 00:15:58.549 --> 00:16:00.889 stock prices? Instead of predicting just the 00:16:00.889 --> 00:16:03.970 next single candle, the idea is the AI generates 00:16:03.970 --> 00:16:06.509 a kind of blurry picture of where the price might 00:16:06.509 --> 00:16:10.009 go over, say, the next 50 candles. A blurry forecast. 00:16:10.169 --> 00:16:12.870 Yeah, initially noisy and uncertain. But then 00:16:12.870 --> 00:16:15.769 the AI iteratively refines that picture, removes 00:16:15.769 --> 00:16:18.409 the statistical noise step by step based on its 00:16:18.409 --> 00:16:20.870 training until the forecast for those 50 candles 00:16:20.870 --> 00:16:23.409 becomes much sharper, more detailed, and more 00:16:23.409 --> 00:16:25.840 confident. So it's building a multi -step forecast 00:16:25.840 --> 00:16:28.360 by refining it over and over. That's the concept. 00:16:29.159 --> 00:16:30.940 It's a different way to think about time series 00:16:30.940 --> 00:16:33.620 prediction, moving beyond just one step at a 00:16:33.620 --> 00:16:36.019 time. Very cutting edge. And what about the kind 00:16:36.019 --> 00:16:37.779 of prediction? Right now it's predicting a single 00:16:37.779 --> 00:16:40.820 price number, right? Like $40 ,150. Correct. 00:16:41.720 --> 00:16:43.799 The current models usually give a point estimate. 00:16:43.980 --> 00:16:46.779 But the real holy grail, the ultimate goal for 00:16:46.779 --> 00:16:49.200 sophisticated trading and risk management, is 00:16:49.200 --> 00:16:52.179 moving towards predicting probabilities. Probabilities, 00:16:52.600 --> 00:16:54.899 meaning what? Instead of just guessing $40 ,000, 00:16:54.899 --> 00:16:58.419 $150, the AI would tell you something like, there's 00:16:58.419 --> 00:17:01.379 a 75 % probability the price will be higher than 00:17:01.379 --> 00:17:03.720 the current price in the next five minutes, a 00:17:03.720 --> 00:17:06.460 20 % chance it stays roughly the same, and a 00:17:06.460 --> 00:17:09.759 5 % chance it goes down. Okay, so it gives you 00:17:09.759 --> 00:17:12.319 odds, not just a single guess. Exactly. That 00:17:12.319 --> 00:17:14.430 allows for much smarter decision -making. You're 00:17:14.430 --> 00:17:16.589 not just betting on a single number being right, 00:17:16.970 --> 00:17:19.430 you're managing risk based on likelihoods. It 00:17:19.430 --> 00:17:21.630 shifts the game from pure prediction towards 00:17:21.630 --> 00:17:24.009 sophisticated risk management. And why is pursuing 00:17:24.009 --> 00:17:26.490 those real probabilities considered the ultimate 00:17:26.490 --> 00:17:28.750 goal for advanced trading? Because it allows 00:17:28.750 --> 00:17:31.029 for that sophisticated risk management, moving 00:17:31.029 --> 00:17:33.890 way beyond just a simple directional bet. You 00:17:33.890 --> 00:17:36.789 can weigh your decisions Much more intelligently. 00:17:37.069 --> 00:17:39.569 Precisely. How much capital do you risk if there's 00:17:39.569 --> 00:17:43.089 only a 60 % chance versus an 85 % chance? It 00:17:43.089 --> 00:17:44.789 changes everything. OK, let's try to pull this 00:17:44.789 --> 00:17:47.450 all together then. If you had to synthesize the 00:17:47.450 --> 00:17:49.269 big takeaway from the source material for our 00:17:49.269 --> 00:17:52.190 listeners. I think it's this. Trying to predict 00:17:52.190 --> 00:17:55.089 the long -term market, the weeks and months ahead, 00:17:55.650 --> 00:17:57.849 that's incredibly difficult, maybe impossible, 00:17:58.150 --> 00:18:00.170 because it's dominated by unpredictable news, 00:18:00.410 --> 00:18:03.549 human psychology, big external factors. It borders 00:18:03.549 --> 00:18:07.109 on gambling. Right. Too many unknowns. But predicting 00:18:07.109 --> 00:18:09.950 the very short term, those tiny one minute moves 00:18:09.950 --> 00:18:12.390 driven by the structure of the order book, the 00:18:12.390 --> 00:18:14.809 mathematical gaps, the sources argue that is 00:18:14.809 --> 00:18:17.150 actually a solvable problem. It's a mathematical 00:18:17.150 --> 00:18:19.329 puzzle, not a psychological one. So the AI isn't 00:18:19.329 --> 00:18:21.809 some magic crystal ball. Not at all. It's a powerful 00:18:21.809 --> 00:18:24.990 microscope. It's a tool that can perceive and 00:18:24.990 --> 00:18:27.230 react to these mathematical patterns in the order 00:18:27.230 --> 00:18:30.109 book that are happening way too fast or way too 00:18:30.109 --> 00:18:32.910 subtle for any human trader to consistently see 00:18:32.910 --> 00:18:35.849 and act on. The opportunity is hidden in the 00:18:35.849 --> 00:18:38.650 math, not in the news headlines or the fear and 00:18:38.650 --> 00:18:40.990 greed index. That seems to be the core message. 00:18:41.250 --> 00:18:44.150 And the final thought here. This AI revolution 00:18:44.150 --> 00:18:47.349 in trading, it's not some sci -fi future. The 00:18:47.349 --> 00:18:49.309 sources make it clear it's happening right now. 00:18:49.630 --> 00:18:52.670 Absolutely. The tools are here. The interesting 00:18:52.670 --> 00:18:55.390 thing is the main challenge maybe isn't even 00:18:55.390 --> 00:18:58.150 writing the code anymore, thanks to these AI 00:18:58.150 --> 00:19:00.940 assistants. So what is the challenge? It's forcing 00:19:00.940 --> 00:19:03.940 yourself to understand those underlying mathematical 00:19:03.940 --> 00:19:06.559 mechanics of the market. The order book dynamics. 00:19:06.759 --> 00:19:09.000 The stuff most people just gloss over while they're 00:19:09.000 --> 00:19:11.180 watching the ticker tape or reading news articles. 00:19:11.480 --> 00:19:13.420 That's where the edge seems to be. So the call 00:19:13.420 --> 00:19:17.279 to action is what? Don't be intimidated. Use 00:19:17.279 --> 00:19:20.220 the tools. Try asking an AI assistant to explain 00:19:20.220 --> 00:19:22.700 a concept or write some basic code to load some 00:19:22.700 --> 00:19:25.779 historical price data. Start exploring. The real 00:19:25.779 --> 00:19:27.940 opportunity seems to be in understanding and 00:19:27.940 --> 00:19:30.119 leveraging the market's hidden math, and the 00:19:30.119 --> 00:19:32.299 tools to do that are finally within reach for 00:19:32.299 --> 00:19:34.940 almost anyone. Focus on the mechanics, use the 00:19:34.940 --> 00:19:37.299 assistance, and see what you can discover. Exactly. 00:19:37.859 --> 00:19:38.720 Out to Ro music.