Deep Learning for Market Data

Deep learning is the technology that taught computers to see, to transcribe speech, and to write fluent prose — a string of victories so total that “just throw a neural net at it” became a reasonable default in half the world’s data problems. Then it walked into financial markets and got quietly, repeatedly humiliated. The reason is everything you learned in Machine Learning for Alpha, turned up to eleven: the signal-to-noise ratio is brutal, the data-generating process drifts under your feet, your samples overlap and bleed information, and now you’ve added a model class with millions of parameters and an almost supernatural gift for memorizing noise. A transformer that aced a billion words of text will, on a few thousand non-stationary return observations, fit the wiggles perfectly and predict the future not at all.

This course is the unhyped version. It is not a victory lap for neural networks in trading — it is an honest field manual for when they help, when they hurt, and how they actually work on the hardest data there is. We assume you’ve internalized the overfitting discipline from Machine Learning for Alpha and the stationarity, autocorrelation and volatility-clustering facts from Time-Series Finance; here we add the deep-learning layer on top, one architecture at a time, always asking the same skeptical question: is this earning its keep, or just its parameter count? The arc:

• Why deep learning struggles in finance — the effective-sample-size catastrophe, why a plain multilayer perceptron both over- and under-fits low-signal returns, and the bias–variance reframing that explains why “bigger model” is usually the wrong reflex.
• Recurrent models: RNN, LSTM & GRU — reading a sequence one step at a time, the unrolling mental model, the vanishing-gradient problem that cripples plain RNNs, and how the LSTM/GRU gating highway lets information survive a long window.
• Temporal convolutional networks — treating a return series like a 1-D signal: causal, dilated convolutions that see far back with few parameters, why they often beat RNNs on tabular financial sequences, and the receptive-field arithmetic.
• Attention & transformers for markets — self-attention as “who looks at whom,” why a direct one-hop connection across time beats a decaying hidden state, transformers on returns and on limit-order-book microstructure (the DeepLOB lineage), and the data-hunger that punishes them on small samples.
• Embeddings for categorical & alternative data — turning tickers, sectors, regimes, calendars and news into dense learned vectors, why entity embeddings beat one-hot encodings, and the leakage traps that embeddings make especially easy to fall into.
• Training & evaluating without fooling yourself — aggressive regularization (dropout, weight decay, early stopping, augmentation) for a tiny sample, purged and embargoed evaluation applied to deep nets, the deflated Sharpe for an architecture search, and the sober bottom line: gradient-boosted trees are the baseline to beat, and usually they win.

By the end you’ll be able to build any of these models — and, more valuably, to know when not to, and to defend the choice with the effective sample size, the receptive field, and a deflated out-of-sample number rather than a glittering backtest. A graded final exam runs the whole discipline back at you at the end, one locked question at a time.