Posted on March 2, 2024
Developing a trading bot, progressing from basic AI concepts to advanced applications in financial analytics. Despite challenges, they experiment with various techniques and models, demonstrating a journey of learning, adaptation, and persistent exploration in AI.
My fascination with artificial intelligence has always been more than a passing interest. I've vigilantly followed the field's evolution, dabbled in crafting some basic neural networks, and felt reasonably confident in my understanding of their mechanics. However, the advent of transformer models and the rise of ChatGPT marked a turning point, signaling that it was time for me to recalibrate my approach and deepen my technical repertoire.
Embarking on this new chapter, my ambitions were loftier than ever. I envisioned creating a trading bot not just functional but also profitable, designed to operate seamlessly on a standard mid-range laptop. This bot would be a tapestry woven from various AI methodologies—Autoencoders to distill immense data pools into essence, Transformers to decode and learn from this essence through supervised training on nuanced buy/sell signals, and Reinforcement Learning to steer the helm, evolving through its interactions.
So, here's to the inception of a thrilling endeavor!
With a roadmap in hand, pinpointing the technologies I needed to master, I initiated my journey by setting up a digital sanctum aptly named "Percepticon." My first foray? A basic perceptron—a humble nod to the genesis of AI learning, akin to scripting a "Hello World."
The training ground chosen was the MNIST dataset, an assembly of 70,000 images, typically split into 60,000 training images and 10,000 test images 28 x 28 pixel grayscale images portraying hand-scribbled numerals. It's a quintessential starting block in AI practice.
My neural network was straightforward, embodying the classic architecture: an input layer with 784 nodes (one for each pixel), a hidden layer hosting 128 nodes, and an output layer designed to identify digits 0 through 9, encoded distinctly. Because of my aversion to the nitty-gritty of matrix math, I conceded to using numpy for sanity's sake.
The process was methodical: flatten the image arrays into 784-element vectors, initialize weights randomly, forward propagate employing ReLU activations, culminating in a Softmax for probability distribution, and then fine-tune via backpropagation until the improvements plateaued. Impressively, the network reached a 98% accuracy rate—a testament to the efficacy of even basic structures in AI.
To put theory into tangible proof, I crafted a rudimentary (undeniably clunky) drawing application. My creation recognized the digits I sketched with a success rate that, though not flawless, demonstrated a clear grasp of the learned patterns—functioning accurately around 80% of the time.
This venture was not just about proving a concept; it was a hands-on affirmation of the foundational knowledge I possessed and a stepping stone toward the more complex landscapes of AI I aimed to navigate.
As I ventured further into the PyTorch landscape, the grand-daddy of neural network libraries, I decided to elevate my project by transitioning from a simple perceptron to an autoencoder. This wasn't just about transformation; it was about truly grasping the intricacies of the autoencoder architecture. Once I felt confident in my understanding, I embarked on the PyTorch translation of this model, aiming for something more than just a basic rendition.
The challenge I chose was ambitious: to develop an autoencoder that could adjust image exposure and fill in under- or overexposed areas—a concept that had intrigued me for years. Why this functionality isn't a staple in advanced image editing tools remains a mystery to me.
My strategy involved using convolutional layers with ReLU activations, compressing the data down to a 32 x 32 latent representation before expanding back to a 256 x 256 image, concluding with a hyperbolic tangent activation, which pushes the values to produce a one or a negative one, to accommodate the -1 to 1 normalization.
After prepping a dataset of 132,354 images, I embarked on a rigorous training regimen. I manipulated the brightness of each 256 x 256 image to craft my training set, initiating an arduous cycle of forward and backward passes that lasted three days on my modest Ryzen 5 laptop—sans the luxury of CUDA acceleration.
The result? A blend of success and frustration. While the autoencoder managed to adjust exposure, the output images were disappointingly blurry, and the anticipated clipped infill remained elusive. Despite tweaking loss functions, network depth, and bottleneck width, sharpness and detail in the reconstructed images remained just out of reach.
Yet, this endeavor wasn't in vain. I emerged with a solid grasp of autoencoders, even if the dream of perfect automatic exposure correction requires further pursuit. For now, this project may be shelved, but it's poised for revival once my laptop transforms from a learning tool into a profit engine.
Embarking on a quest to harness stock data for predictive modeling, I dove headfirst into the tumultuous world of financial analytics. Selecting a dozen prominent large-cap stocks, I aimed to construct a neural network capable of forecasting the subsequent day's stock candle, drawing on a year's worth of Open, High, Low, Close, and Volume data.
Normalizing data in the realm of imagery is straightforward, given the bounded nature of pixel values. However, stock data presents a wilder frontier, devoid of inherent caps. What is the ceiling for a stock's value? My solution was to employ ratios relative to the previous close, a method I'll refine in future iterations to offer more granularity and relevance.
The normalization process, particularly the transformation of these ratios to a 0-1 scale, now strikes me as potentially flawed—a misstep in my haste that might have muddled the model's interpretations.
I allocated a decade of data, reserving the last year for testing after nine years of training. What followed was a relentless debugging marathon, fraught with errors and makeshift fixes. The experience was akin to navigating a minefield, fraught with setbacks yet punctuated by glimpses of progress.
Although the model exhibited a semblance of functionality, accurately predicting the general trend and magnitude of stock movements, the accuracy of price levels remained elusive. Whether due to flawed denormalization or minor computational discrepancies, the precise scaling of predicted values proved challenging. Despite these hurdles, the model's directional predictions resonated with uncanny accuracy across numerous trials.
In retrospect, my perception of success may be tinted by optimism, a willful search for validation amidst the chaos. Yet, in this crucible of trial and error, I found invaluable lessons and a renewed resolve to refine my approach in future endeavors.
Determined to rectify past missteps and pursue precision, I embarked on a new journey, this time armed with a clear vision and a commitment to methodical execution. The objective was straightforward yet ambitious: to leverage the power of Reinforcement Learning through Stable Baselines Python library, integrating a diverse array of data inputs to forge a sophisticated predictive model.
Transitioning from isolated datasets to a dynamic influx of information, I turned to yfinance, a library to access Yahoo Finance, for real-time data acquisition, enriching my analysis with an exhaustive suite of indicators via a technical analysis library called pandas_ta. Inspired by a novel approach I stumbled upon—a YouTuber's strategy employing K-Nearest Neighbors for peak and trough detection—I decided to experiment with this technique, intrigued by its simplicity and apparent efficacy. K-Nearest Neighbors (KNN) clusters and categorize data, one of these things is just like these others.
Upon implementing the KNN model, I observed a pattern: while false positives were present, they predominantly clustered around actual market peaks and troughs, suggesting a nuanced, albeit imperfect, discernment of critical trends. Encouraged by these preliminary signs, I ventured into backtesting, seeking to validate the model's practical value.
The results were a mixed bag—though the strategy did not consistently outperform a straightforward buy-and-hold approach, it demonstrated sporadic bouts of superior profitability. This outcome, while not unequivocally victorious, hinted at the potential lurking within these analytical methodologies.
In an unforeseen twist, my focus veered toward an unexpected ally: the Random Forest algorithm. A random forest creates many decision trees using different parts of the data and combines their results to make more accurate predictions. Admittedly, my grasp of its inner workings—echoes of Markov chains and statistical averages—was less than complete. Yet, its reputation for versatility spurred me to integrate it into my analysis, smoothing data with a centered window average to discern peaks and valleys for training the model.
Astoundingly, this approach demonstrated an ability to outperform both standard buy-and-hold strategies and the broader S&P 500 across numerous assets. This accidental discovery propelled me down an exploratory detour, brimming with potential yet fraught with complexity.
My on to a forward-testing phase, conducted through paper trading on daily intervals. The model has shown prowess within lateral market phases. However, its performance faltered amidst pronounced bullish or bearish trends, prompting a quest for adaptive mechanisms to discern and react to market transitions effectively.
Experimentation with supertrend adjustments and support/resistance analysis offered incremental advancements but fell short of transformative impact. Meanwhile, the idea of calibrating predictive accuracy based on recent performance metrics lingered as a tantalizing prospect.
As I oscillate between my primary Reinforcement Learning project and these unforeseen quantitative diversions, the allure of optimization remains an irresistible siren call. Though initially conceived as a mere analytical tool, this Random Forest model has burgeoned into a standalone quantitative strategy, its true efficacy yet to be fully ascertained.
In the interim, as I await the empirical verdict from ongoing forward tests and refine my overarching methodologies, the boundary between focused research and captivating distraction blurs. Yet, within this maelic of innovation, the journey itself emerges as the most enlightening reward.