FX Spot Pairs
Momentum and carry factors in the world's most liquid market
Foreign Exchange
Daily
Price Data
Methodology Highlight
Teaches horizon-sensitivity analysis and hypothesis revision — the same features behave differently across return horizons, requiring systematic evaluation rather than single-configuration testing.
This case study applies the ML4T workflow to 20 G10 currency pairs using daily data from OANDA. Foreign exchange presents a structurally challenging prediction problem: the cross-section is small (20 pairs dominated by a single USD factor), limiting diversification and effective breadth.
Students learn to build carry and momentum features for FX, and discover how prediction horizon matters — the same features can be informative at one horizon and uninformative at another. The case study teaches horizon-sensitivity analysis as a systematic diagnostic step.
The pipeline progresses from momentum to carry to combined signals, illustrating hypothesis revision as a natural part of the research process. Students also learn how to apply causal inference (DML) to assess confounding in momentum signals, and how to set pair-specific cost budgets in a market where execution costs vary significantly across pairs.
Strategy Summary
Long-short daily strategy on 20 G10 FX pairs ranked by momentum and carry signals. Daily NY 5PM close cadence with next-bar-open execution. Dollar- neutral constraints. The case study evaluates multiple return horizons (1-day, 5-day, 21-day) to understand horizon sensitivity. 8 CV folds with 5-year training and 1-year validation windows.
Data Sources
OANDA (spot rates)
FRED (interest rate differentials)
ML Techniques
Carry factor construction
Momentum signals across horizons
Causal DML for confounding
Horizon-sensitivity analysis
ML Pipeline
Universe & Setup
20 G10 currency pairs from OANDA. Daily NY 5PM close cadence. Carry decomposition
(spot return + interest rate differential) tracked from setup. Dollar-neutral
constraints with ~3 effective independent bets (USD dominates the cross-section).
8 CV folds (5Y train, 1Y val) with 2024-2025 holdout. Cost model: 1-3 bps for
majors, 3-8 bps for crosses.
Universe & Protocol Setup
Ch 6
Defines the G10 FX universe with NY 5PM close convention, carry
decomposition (spot return vs interest rate differential), and dollar-neutral
constraints. Performs horizon feasibility analysis comparing 4-hour and daily
bar moves against typical spreads. Builds walk-forward evaluation splits
with purge/embargo rules.
Labels & Evaluation
Three forward return horizons (1-day, 5-day, 21-day) reveal a ninefold IC scaling
from 1d to 21d -- the cross-sectional momentum effect operates at monthly, not daily
frequency. Evaluation covers 61 features (51 financial + 10 temporal) with HAC-adjusted
IC. The horizon discovery drives the subsequent modeling to focus on the 21-day label.
Label Engineering
Ch 7
Aggregates 4-hour bars to daily using NY 5PM rollover convention and computes
forward return labels at three horizons (1-day, 5-day, 21-day). Applies
session-bounded forward returns to prevent weekend gap leakage. Builds CV
configuration respecting the FX calendar with appropriate purge windows per
horizon.
Feature Evaluation
Ch 7
Evaluates all features (financial + temporal) against the 1-day forward
return label using HAC-adjusted IC with fold-level stability analysis.
Applies Benjamini-Hochberg FDR correction and screens for
coverage and staleness. Triages features into PROCEED / REVISE / STOP
categories for downstream modeling.
Feature Engineering
51 financial features spanning mean-reversion (z-scores, distance from extremes,
OU half-life), multi-horizon momentum, risk-adjusted returns, and carry metrics.
10 temporal features from local linear trend Kalman filter with MLE noise estimation,
HMM regime detection, and spectral analysis. 61 features total. The carry decomposition
from setup feeds directly into carry-specific features unique to FX.
Feature Engineering
Ch 8
Builds features spanning mean-reversion (z-scores, distance from extremes,
OU half-life), multi-horizon momentum, risk-adjusted momentum, volatility
(Garman-Klass at multiple windows), carry proxies, and cross-sectional rank
features across the pair universe. Adds USD factor and rate-differential
proxies unique to the FX setting.
Temporal Features
Ch 9
Fits walk-forward temporal models: local linear trend
Kalman filter per pair with MLE noise estimation, HMM for USD volatility
regime detection on aggregate currency movements, and ARIMA residual features.
All models are fitted on training windows and applied forward without
re-estimation.
Modeling
GBM leads at IC +0.045 on the 21-day label. Deep learning produces 0% positive
backtests -- the starkest DL-negative result across all nine case studies, likely
because the 20-pair cross-section is too small for attention or convolution to add
value. DML reveals 60% confounding bias -- naive momentum overstates the true effect
by more than half.
Linear Models
Ch 11
Trains Ridge, LASSO, and ElasticNet via walk-forward CV on the FX pair
universe with all features. Generates predictions across all folds and
evaluates cross-sectional IC per configuration. Compares regularization
approaches and stores coefficients for feature importance analysis.
Gradient Boosting
Ch 12
Trains LightGBM across regularization profiles and loss functions with IC
evaluated at iteration checkpoints. Tests whether shallow trees with heavy
regularization outperform deeper configurations given the narrow FX
cross-section. Registers all predictions for downstream backtesting.
Tabular Deep Learning (TabM)
Ch 12
Trains TabM rank-1 adapter MLP ensembles (small/medium/large configurations)
via walk-forward CV on the FX feature matrix. Compares attention-based
cross-pair interaction learning against GBM and linear baselines. Registers
predictions for downstream backtesting.
TCN
Ch 13
Trains dilated causal TCN on 60-day lookback windows of the FX feature
panel. Tests whether temporal evolution of features captures signal that
flat models miss. Compares IC learning curves against linear, GBM, and TabM
baselines.
NLinear
Ch 13
Trains NLinear (last-value subtraction plus single linear layer) as the
minimal temporal DL baseline for FX. Tests the Zeng et al. (2023) hypothesis
that simple linear architectures match complex ones on time series forecasting.
Compares against TCN, TabM, GBM, and linear results.
Causal DML
Ch 15
Applies DML to skip-recent momentum (mom_skip_recent) as treatment with
three confounders -- Garman-Klass volatility at 21-day and 63-day horizons
plus 21-day z-score -- across G10 FX pairs. Finds substantial amplification
from naive to DML effect with significant confounding bias. Runs
block-permutation placebo refutation confirming a genuine causal reversal
consistent with the overshooting hypothesis.
Strategy Pipeline
Only holdout that improved: validation Sharpe +0.19 to holdout +0.28 (+45%), likely
reflecting a favorable carry regime. Strategy turns negative above 15 bps per leg --
the tightest cost tolerance with the narrowest viable operating window. Pair-specific
cost budgets needed rather than blanket assumptions. Position-level exit rules tested.
Model Analysis
Ch 11
Compares all model families (linear, GBM, TabM, DL, causal) on the FX
prediction task using registry metrics, fold stability diagnostics, feature
importance, prediction correlation, and regime-conditional IC analysis.
Produces per-family advancement recommendations for Ch16 backtesting.
Backtest & Signal Evaluation
Ch 16
Runs plumbing test (random signal verification), then sweeps all model
predictions across signal methods and TOP_K configurations using the
ml4t-backtest engine with daily FX prices. Computes DSR, family-level
comparison, and IC-to-Sharpe translation statistics.
Portfolio: Allocator Sweep
Ch 17
Sweeps top signal-stage predictions across TOP_K concentration levels and
6 allocators (equal-weight, score-weighted, inverse-vol, risk-parity, MVO,
HRP) on the FX pair universe. Tests how concentration interacts with
allocation method in a small cross-section with few independent risk factors.
Transaction Costs
Ch 18
Runs a cost grid sweep on top allocation-stage combinations to find the
breakeven cost level. Plots net Sharpe decay curves and identifies the viable
cost window for FX majors. Compares breakeven against
realistic interbank execution costs.
Risk Management
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.
Calibrates trailing stops using MAE/MFE distributions. Measures how each
overlay modifies drawdown and Sharpe without adding excessive turnover.
Synthesis & Verdict
Genuine signal in the most efficient market on earth, but fragile. Verdict: Advance --
build pair-specific cost models from actual interbank spreads. The 21-day horizon is
the only viable operating point. A story of hypothesis revision from momentum to carry
to the conclusion that cost precision matters more than model complexity.
Strategy Analysis
Ch 20
Assembles the full FX pipeline verdict by tracing the champion through
signal, allocation, cost, and risk stages via BacktestExplorer. Computes
holdout performance, search risk accounting, and operating profile. Produces
a structured iterate/advance verdict with factor attribution for Ch20
synthesis.
Quick Info
Asset Class
Foreign Exchange
Frequency
Daily
Data Type
Price Data
Notebooks
17
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