NASDAQ-100 Microstructure
Intraday microstructure signals across 114 stocks at 15-minute frequency
Equities
15-Minute
Microstructure
Methodology Highlight
Shows how rebalance frequency and cost regime — not model quality — determine whether a statistically significant signal translates into a viable strategy.
This is the highest-frequency case study in the book, using AlgoSeek TAQ-derived 15-minute bars for 114 NASDAQ-100 constituents. Students learn to build features from order flow, quote staleness, relative spreads, and other microstructure indicators — the richest feature space in the book with 88 features from 4 microstructure families.
The case study demonstrates how rebalance frequency and transaction costs interact to determine strategy viability. The same prediction signal can be worthless or profitable depending on how frequently positions are adjusted and what cost model is applied. Students learn to use per-share cost models rather than basis-point approximations for intraday strategies.
With the largest training set in the book (13M+ observations across 114 stocks), this case study also illustrates the practical challenges of working with high-frequency data: storage, computation time, and the difference between statistical significance and economic significance.
Strategy Summary
Intraday rank-and-trade across 114 NASDAQ-100 stocks using 15-minute bars. Dollar-neutral long-short with equal-weight sizing within legs. Cost model includes spread, market impact, and commission components. Walk-forward evaluation uses 2 folds over 2020-2021. The critical design choice is rebalance cadence — the strategy is evaluated at multiple frequencies to understand the cost-signal tradeoff.
Data Sources
AlgoSeek (TAQ-derived bars)
ML Techniques
Order flow features
Microstructure indicators
Classification vs regression labels
GBM with large-scale data
ML Pipeline
Universe & Setup
114 NASDAQ-100 constituents from AlgoSeek TAQ minute bars, aggregated to 15-minute
intervals. Highest-frequency case in the book. Dollar-neutral constraint, 5+ bps
friction floor that dominates the cost model. Only 2 CV folds (6-month train/val)
due to limited 2020-2021 history. Execution delay is 1 bar (15 minutes).
Universe & Protocol Setup
Ch 6
Defines the NASDAQ-100 universe with 15-minute bar aggregation,
close definition (last trade in bar), 1-bar execution delay, and next-bar-open
fill assumption. Performs friction hurdle analysis comparing per-bar return
distributions against typical large-cap spreads. Builds 2-fold walk-forward
splits with purge/embargo calibrated for 15-minute horizon.
Labels & Evaluation
15-minute forward return as primary label using midprice (not last trade) to reduce
microstructure noise. 5-minute and 60-minute variants for cadence analysis. 88
features (66 financial + 22 temporal) evaluated. Classification labels dominate
regression -- a pattern first discovered here and repeated across case studies.
Label Engineering
Ch 7
Computes midprice-based forward returns at 15-minute (primary), 5-minute, and
60-minute horizons using NBBO quotes to avoid bid-ask bounce contamination.
Applies session-bounded forward returns to prevent overnight gap leakage.
Builds a ternary classification label (fwd_dir_15m) with a threshold and
generates walk-forward CV configuration.
Feature Evaluation & Triage
Ch 7
Evaluates all features (financial + temporal) against 15-minute forward
midprice returns using HAC-adjusted IC (session-length Newey-West for
overlap) with BH-FDR correction. Screens for coverage and staleness, assesses
quantile monotonicity and cross-feature redundancy. Triages features into
PROCEED / REVISE / STOP categories.
Feature Engineering
66 microstructure features across four families: quote-based liquidity (midprice,
spread dynamics, depth imbalance), trade-based flow (signed volume share, trade
intensity), intraday volatility (Garman-Klass, realized variance), and regime
proxies. 22 temporal features including HAR(5,15,60) volatility decomposition unique
to intraday data. The richest per-bar feature set in the book.
Microstructure Feature Engineering
Ch 8
Engineers features across four microstructure families from AlgoSeek minute
bars: quote-based liquidity (midprice, spread, depth imbalance), order flow
proxies (signed volume, tick imbalance, microprice deviation),
volatility/impact (realized vol, Kyle's lambda, Amihud), and hidden liquidity
(FINRA dark pool share). Computes multi-resolution features at 1-bar, 5-bar,
15-bar, and 60-bar lookbacks with 5-bar staleness caps on quote-derived fields.
Temporal Features
Ch 9
Constructs temporal features using three approaches tailored to intraday
data: HAR(5,15,60) volatility decomposition for multi-scale realized
volatility, rolling FFT spectral features on volume and volatility profiles,
and depth-2 path signatures on (price, signed_vol, trades) trajectories using
causal rolling windows.
Modeling
4 deep learning architectures tested (most of any case study): NLinear as minimal
baseline, LSTM for order-flow memory, TCN for dilated causal convolutions, PatchTST
for multi-scale attention. GBM trains on 13M+ samples (largest training set). Best
IC only +0.008 (weakest in book). Causal DML tests whether signed volume share
causes 15-minute returns.
Linear Models
Ch 11
Trains Ridge, LASSO, and ElasticNet via walk-forward CV on microstructure
features across NASDAQ-100 symbols at 15-minute frequency.
Establishes the linear IC baseline for downstream model comparison and
registers predictions for backtesting.
Gradient Boosting
Ch 12
Trains LightGBM at 15-minute frequency across regularization
profiles and loss functions (MSE, MAE, Huber). Evaluates IC at 50-iteration
checkpoints and compares against the linear baseline.
NLinear (DL Baseline)
Ch 13
Trains NLinear (last-value normalization plus single linear map) as the
minimal temporal DL baseline for the NASDAQ-100 microstructure case.
Establishes clean provenance for architecture comparison against LSTM, TCN,
and PatchTST.
LSTM
Ch 13
Trains LSTM with gated memory on 60-bar lookback windows of microstructure
features at 15-minute frequency. Tests whether recurrent memory captures
short-lived order-flow and spread dynamics beyond memoryless models. Compares
against NLinear, Ridge, and GBM results.
TCN
Ch 13
Trains dilated causal TCN on 15-minute microstructure features. Tests whether
multi-scale convolutional receptive fields capture intraday temporal dynamics
that recurrent memory misses. Compares against LSTM, NLinear, and prior
baselines on the same 15-minute label.
PatchTST
Ch 13
Trains PatchTST with multi-scale patch attention on 15-minute bar sequences.
Tests whether patching plus self-attention captures the microstructure edge
more effectively than NLinear, LSTM, and TCN. Produces the final DL candidate
for downstream model analysis and backtesting.
Causal DML Estimation
Ch 15
Applies DML to signed volume share (buyer-initiated volume fraction via
Lee-Ready) as treatment across NASDAQ-100 stocks at 15-minute frequency, with
confounders: relative spread, 5-minute realized volatility, and 1-month
cumulative return. Runs placebo refutation tests and quantifies confounding
bias.
Strategy Pipeline
Despite the weakest IC, produces the highest Sharpe (+4.22) through frequency
multiplication: 26 daily decisions x 114 stocks = massive effective breadth. The
same signal is worthless at per-bar frequency and profitable at hourly. Per-share
cost models (not basis-point sweeps) reveal the relevant budget. Flagship cost
analysis notebook.
Cross-Model Analysis
Ch 11
Compares all model families (linear, GBM, DL, causal) on the highest-frequency
case study using registry metrics, fold stability diagnostics, prediction
bucket monotonicity, and regime-conditional IC. Produces per-family advancement
recommendations for Ch16 backtesting across 2 walk-forward folds.
Signal-Stage Backtest
Ch 16
Runs plumbing test (random signal verification), then sweeps all predictions
across signal methods using the ml4t-backtest engine with 15-minute OHLCV bars
from AlgoSeek TAQ. Computes DSR, family comparison, and IC-to-Sharpe
translation statistics. Registers results for downstream allocation and cost
analysis.
Portfolio Allocator Sweep
Ch 17
Sweeps top signal-stage predictions across TOP_K concentration levels and
allocators (equal-weight, score-weighted, inverse-vol, risk-parity) for the
NASDAQ-100 universe. Skips MVO/HRP due to computational cost at 15-minute
cadence. Dollar-neutral constraint throughout.
Transaction Cost Analysis (Cadence Sweep)
Ch 18
Runs two cost analyses: a standard bps cost grid on top allocation combos,
and a cadence-by-per-share cost sweep that varies rebalance frequency from
15-minute to 4-hourly. Uses a per-share cost model ($/share) rather than bps,
which is more realistic for equities. Identifies the viable implementation
regime where signal survives execution friction.
Risk Controls
Ch 19
Sweeps position-level (stop-loss, trailing stop, time exit) and portfolio-level
(drawdown breaker, daily loss limit) risk controls on top allocation combos
with 15-minute OHLCV bars. Measures how each overlay modifies the equity curve
and drawdown profile at intraday granularity.
Synthesis & Verdict
Weakest IC, highest Sharpe -- the Fundamental Law of Active Management in action.
Holdout Sharpe +2.05 (51% decay, still strong). Verdict: Advance, but cadence
optimization is the priority, not model improvement.
Strategy Synthesis & Verdict
Ch 20
Assembles the full NASDAQ-100 pipeline verdict by tracing the DL regression
champion through signal, allocation, cost, and risk stages via
BacktestExplorer. Computes holdout performance, search risk accounting, and
factor attribution. Produces a structured deployment verdict with cadence
optimization as binding constraint.
Quick Info
Asset Class
Equities
Frequency
15-Minute
Data Type
Microstructure
Notebooks
18
Chapters
13
Chapters
6
Strategy Research Framework
7
Defining the Learning Task
8
Financial Feature Engineering
9
Model-Based Feature Extraction
11
The ML Pipeline
12
Advanced Models for Tabular Data
13
Deep Learning for Time Series
15
Causal Machine Learning
16
Strategy Simulation
17
Portfolio Construction
18
Transaction Costs
19
Risk Management
20
Strategy Synthesis