Deep Learning for Market Data — Final Exam

This is the graded finale for Deep Learning for Market Data, and it runs across everything the course built. You started by confronting why deep nets struggle on price data at all — the effective sample size is tiny once you account for autocorrelation, so a flexible MLP is forever squeezed between over- and under-fitting, and the universal-approximation theorem is no rescue when every extra trial taxes your significance. From there you walked the architecture ladder: recurrent nets (RNN/LSTM/GRU) that unroll through time but bleed gradients ( $0.6^{20}\approx 3.6\times10^{-5}$ ) until the LSTM’s cell-state highway and gates fix the leak; temporal convolutional networks whose causal, dilated convolutions guarantee no look-ahead and grow a receptive field $R = 1 + (k-1)(2^{L}-1)$ in parallel; and attention/transformers, whose scaled dot-product ( $\sqrt{d_k}$ ) gives any two timesteps a one-hop path at $O(T^2)$ cost, powering LOB/DeepLOB models on the high-SNR order book. You learned to turn sparse categoricals into dense entity embeddings without leaking the future, and finally to train and evaluate honestly — the regularization menu, purged + embargoed cross-validation, the deflated Sharpe hurdle for architecture hunts, the deployment checklist, and the humbling fact that gradient boosting usually wins on tabular market features. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.

Course Recap

Big picture

Deep Learning for Market Data — the whole arc

DL for Market Data
- 1 · Why DL struggles
  - Effective sample size ≪ row count (autocorrelation)
  - MLP squeezed: overfit if flexible, underfit if shrunk
  - Universal approximation ≠ learnable; every trial taxes significance
- 2 · RNN / LSTM / GRU
  - Recurrence = unroll the same cell through time
  - Vanishing/exploding gradient over long horizons
  - LSTM cell-state highway + gates; GRU is the lean cousin
- 3 · Temporal Conv Nets
  - Causal convolution → no look-ahead leakage
  - Dilation grows receptive field R = 1+(k−1)(2^L−1)
  - Residual blocks; parallel training vs sequential RNN
- 4 · Attention & Transformers
  - Scaled dot-product, divide by √(d_k)
  - One-hop path O(1) vs RNN O(T); cost O(T²)
  - Multi-head + positional encoding; LOB/DeepLOB
- 5 · Embeddings
  - One-hot wastes space; embeddings are dense + learned
  - Dim rule of thumb ≈ min(50, cardinality/2)
  - Leakage traps: fit on train only, no future labels
- 6 · Training & evaluation
  - Regularization menu: dropout, weight decay, early stop
  - Purged + embargoed CV; deflated Sharpe for arch search
  - Gradient boosting usually wins on tabular features

Six lessons, one ladder: from why DL struggles on prices to honest evaluation.

One run, one shot

This is a graded, irreversible exam. There are 24 questions, shown one at a time. The instant you submit a question it locks — there is no Back button, no retry, and no Restart. A wrong answer simply fails that question and the exam moves on; you cannot revisit it. Your running score is hidden until the final screen. The pass mark is 70%. Some questions accept more than one correct option — read every option before you commit, because once you submit you own the answer.

Question 1 of 27

A quant has 20 years of daily equity returns — about 5,000 rows — and wants to justify a deep network with 200,000 parameters. Why is the row count a misleading measure of how much data she actually has?

Daily data is always too low-frequency for any neural network regardless of length.Twenty years is plenty; 5,000 rows comfortably support 200,000 parameters.Returns are strongly autocorrelated, so the effective sample size is far smaller than 5,000 independent observations.Parameter count is irrelevant to data needs as long as you use dropout.

Select an answer to continue.

Where this leaves you on the quant ladder

Pass or fail, you now hold the rare skill that separates practitioners from hype-followers: knowing when not to reach for a deep net. You can size up a dataset’s effective sample, pick the architecture whose inductive bias actually fits the problem — gated recurrence, a dilated causal stack, or attention over the order book — wire categoricals in with leakage-safe embeddings, and then defend a result against the deflated-Sharpe hurdle and a purged, embargoed split. That is exactly the judgment that lets you deploy deep learning where it genuinely earns its keep and lean on gradient boosting everywhere else. From here the ladder climbs into full production systems and portfolio construction — and you arrive with the one habit that matters most in quant: trusting evaluation over enthusiasm.