An Investigation of Exploitability
Working Paper DAI-2605 | Dissensus AI
This paper investigates whether distributional asymmetries in foreign exchange alpha signals represent exploitable market inefficiencies. Using EUR/JPY data spanning November 2015--August 2025 (504 weekly observations after rolling window warmup), we find that the asymmetry premise itself largely dissolves under dependence-robust measurement: of five alpha signal types, only the volatility-expansion (coverage) signal exhibits skewness that survives block-bootstrap inference (1.75, 95% CI [1.17, 2.15]); the signed tail signal skews negative rather than positive (-1.47, CI [-3.09, 0.54]), and momentum, mean-reversion, and correlation signals are statistically indistinguishable from symmetry. The economic null is correspondingly stark: a skewness-threshold strategy earns 3.60% cumulative gross over the decade (17 trades, Sharpe 0.15), is nearly inert in walk-forward testing (3 trades in eight out-of-sample years), and carries no residual alpha against proxy carry, momentum, and dollar factors. Because the strategy trades so rarely, transaction costs are immaterial rather than decisive: net returns remain within 0.4 percentage points of gross even at wide retail spreads, with break-even near 19 pips round-trip. Data-snooping corrections complete the null: White's Reality Check (p = 0.15) and Hansen's SPA (p = 0.28) find no strategy in a 13-candidate universe that outperforms -- the best realized mean return belongs to the random benchmark. We conclude that alpha signal asymmetry in this setting is neither robustly detectable nor exploitable, and we caution that unsigned-magnitude constructions can manufacture the appearance of asymmetry where none exists.
| Finding | Result |
|---|---|
| Alpha signals deviate from normality? | Mostly -- 4 of 5 reject; fast alpha is exactly Gaussian |
| Skewness robust to serial dependence? | Only coverage alpha; the tail signal skews negative and fragilely |
| Exploitable heavy tails? | No (GPD shape -0.25, CI [-1.62, 0.25]; weak clustering, extremal index 0.83) |
| Strategy returns distinguishable from zero? | No -- the central null; return and Sharpe CIs include zero before costs |
| Do transaction costs matter? | No -- immaterial at 17 trades/decade (break-even ~19 pips round-trip) |
| Survives data-snooping correction? | No -- RC p = 0.15, SPA p = 0.28; best raw performer is the random benchmark |
| Cross-market generalization? | No -- the tail-skew signature reverses sign in GBP/USD, SPY, and GLD |
This is a null result paper, and the null starts earlier than the usual backtest disappointment: under dependence-robust measurement, most of the claimed asymmetry was never there. An earlier version of this paper reported pronounced positive tail skewness (5.05); that figure described the unsigned exceedance magnitude, which is right-skewed by construction. The corrected paper documents that measurement failure from the inside -- a case study in how higher-moment "stylized facts" can be manufactured by sign conventions. Null findings of this kind are underreported in quantitative finance (Harvey, 2017), yet they prevent wasted research effort and capital allocation to spurious patterns.
| Alpha Type | Description | Skew (signed series) |
|---|---|---|
| Tail Alpha | Signed returns beyond the rolling 95th-percentile magnitude threshold | -1.47 (CI [-3.09, 0.54]) |
| Fast Alpha | 5-day return normalized by 20-day realized volatility | 0.01 |
| Pricing Alpha | Deviation from 60-day fair value (mean reversion) | -0.17 |
| Coverage Alpha | Volatility compression ratio σ₂₀(t)/σ₂₀(t−5) − 1 | 1.75 (CI [1.17, 2.15]) |
| Hedge Alpha | DXY correlation × JPY--USD rate differential | 0.15 |
Skewness computed on the signed weekly series (n = 504) with 95% circular block bootstrap intervals; only coverage alpha's interval excludes zero.
null result, alpha asymmetry, foreign exchange, skewness, market efficiency, extreme value theory
G11, G14, G15, C58
alpha-asymmetry/
├── paper/
│ ├── alpha-asymmetry.tex # LaTeX source
│ ├── alpha-asymmetry.pdf # Compiled paper
│ ├── references.bib # Bibliography
│ └── *.png # Figures
├── analysis/
│ ├── full_pipeline.py # Replication pipeline (all tables & stats)
│ ├── full_pipeline_results.json # Pipeline outputs (machine-readable)
│ ├── full_pipeline_results.txt # Pipeline outputs (human-readable)
│ ├── make_asymmetry_figure.py # Figure 1 script
│ ├── make_backtest_figure.py # Figure 2 script
│ ├── phase0_data_verification.py # Data verification
│ └── recompute_tables.py # Legacy table recomputation
├── CITATION.cff
└── LICENSE
Data are retrieved programmatically from public sources (Yahoo Finance) by analysis/full_pipeline.py, which records retrieval counts and dates.
- v2.1 (July 2026, this repository): corrected analysis -- signed tail-signal series, block-bootstrap skewness inference, recomputed benchmark and cost tables, valid extremal-index interval. Each correction is documented in the corresponding table note.
- v2.0.x (Zenodo/SSRN): pre-correction preprint reporting the unsigned-magnitude tail skew (5.05); superseded by this version. The Zenodo concept DOI resolves to the latest deposited version.
@article{farzulla2026alpha,
author = {Farzulla, Murad},
title = {Alpha Asymmetry in Foreign Exchange Markets: An Investigation of Exploitability},
year = {2026},
journal = {Dissensus AI Working Paper DAI-2605},
doi = {10.5281/zenodo.18638784}
}- Murad Farzulla -- Dissensus AI & King's College London
- ORCID: 0009-0002-7164-8704
- Email: murad@dissensus.ai
- Paper (Zenodo): 10.5281/zenodo.18638784
- Paper (SSRN): SSRN:6147567
- Code (GitHub): github.com/dissensus-ai/alpha-asymmetry
- ASCRI Programme: systems.ac/2/DAI-2605
- Dissensus AI: dissensus.ai
Paper content: CC-BY-4.0