Spectral Signatures in SPY and BTC Intraday Returns

Overview

Implementation

Core libraries

NumPy, pandas, SciPy, PyWavelets, pandas-market-calendars, Matplotlib

Data sources

yfinance 1-minute SPY bars, the Binance public klines endpoint for 1-minute BTCUSDT, and FRED VIXCLS for the stress split

Decomposition settings

30-minute Hann STFT with 15/30/60-minute sensitivity windows; a db4 discrete wavelet to five levels with symmetric, periodic and zero boundary sensitivity

Bands

Four edges in minutes — 1–5, 5–15, 15–60, and 60 and above

Configuration

A single frozen Config dataclass carrying dates, symbols, session hours, transform settings, band edges, stress cutoffs, the KS alpha and its Bonferroni divisor of four, and the run seed

Tests

18 pytest modules, roughly one per source module

Components

src/loaders/

SPY, BTC and VIX acquisition, each writing a hashed parquet snapshot through an atomic temp-file rename

src/alignment.py

Regular-trading-hours calendar construction, matched-session alignment, and the separate 24-hour BTC series

src/returns.py

Log returns and the sparse-session rule that drops days with too few observations

src/stft.py, src/dwt.py

The two decompositions, each returning per-band energy through a common interface

src/aggregation.py

Band shares from raw energies, pooled and rolling

src/bootstrap.py

Date-block resampling, so intraday dependence is not broken by the resampler

src/ks.py

Two-sample Kolmogorov–Smirnov tests at the Bonferroni-corrected alpha

src/stress.py

The stress overlay: a VIX median split for SPY, a realized-volatility quantile for BTC

src/crosscheck.py

STFT-versus-wavelet agreement, reported rather than reconciled

src/hashing.py, src/runtime_metadata.py

SHA-256 file digests and the run manifest recording environment and inputs

src/figures/

Nine figure builders, each reading only from processed artifacts

Pipeline

1. src/loaders/* Fetch and snapshot raw SPY, BTC and VIX, recording a digest for each file.
2. build_processed.py Align to the trading calendar, build log returns, and drop sparse sessions.
3. stft.py / dwt.py Decompose both series under both transforms with the configured sensitivity settings.
4. aggregation.py Reduce coefficient energies to band shares, pooled and rolling.
5. build_ks_results.py Bootstrap the band-share distributions and run the corrected KS comparisons.
6. crosscheck.py Quantify STFT-versus-wavelet disagreement on the same input.
7. figures/ Render the nine report figures from processed artifacts only.
8. build_report.py Assemble the report against the manifest, so a stale input cannot be published silently.

Reproducibility and validation

• Every raw and processed artifact is hashed into a run manifest, and the report builder reads figures and tables through that manifest rather than from the filesystem directly.
• Sensitivity settings are declared in configuration rather than run ad hoc: three STFT window lengths and three wavelet boundary modes are part of the standing result, not a follow-up.
• Parquet writes go to a temporary file and are renamed into place, so an interrupted run cannot leave a half-written input behind.
• The multiple-comparison correction lives in configuration — ks_alpha over a Bonferroni divisor of four, one per band — rather than being applied by hand at reporting time.

Availability