Systematic trading researchTail moves in gold, and what they mean for an averaging grid
Huyen X. Vu · Quantex Research, AlgoTrade Pte. Ltd., Singapore · ORCID 0000-0002-2209-842X
Abstract
We test 11 pre-listed hypotheses on 5,349 trading days of spot gold (2004 to 2025) and 156,502 baskets of a simulated averaging grid, then repeat them on a broker feed from 2022 to October 2026. Days beyond 4 standard deviations are about 59 times more common than a normal law predicts. An extreme day makes another within five days 16 points more likely, and extreme days seldom start from quiet markets; these findings hold when whole months are resampled. The grid's adverse moves are heavy-tailed (tail index 2.9). Deep baskets appear 58% more often when the 20-day ATR is 20 USD or more, but the gap disappears once ATR is scaled by price (rate ratio 1.01, 95% CI 0.88 to 1.17), showing the link is with the price level, not volatility. The average daily range rose from 24 USD in 2022 to 123 USD in 2026, faster than price, though this rests on few yearly points.
Keywords: gold; XAUUSD; fat tails; volatility clustering; tail index; averaging grid; price level; out-of-sample
1. Introduction
Averaging grids are built for ordinary days; their risk sits in the few extreme ones. This report examines how gold's extreme days behave and how they affect a simulated grid.
An averaging grid adds orders as price moves against it and closes the whole basket at a small gain. Its risk sits in the tail: the few days when price keeps running. We test 11 pre-listed statements about those days and about one simulated grid, then repeat the tests on a second feed.
2. Data and method
- Main sample: 6,600,530 one-minute XAUUSD bars, 11 June 2004 to 1 October 2025 (5,349 trading days), from the public Kaggle dataset by novandraanugrah (CC0), as copied on Hugging Face (ZombitX64, revision ff6469dc). The feed time zone is not stated; about 121 days of 2025 are missing.
- Second feed: 1,560,327 one-minute bars from one retail broker, 16 May 2022 to 9 October 2026 (1,137 days). Only October 2025 to October 2026 is new time; the rest overlaps the main sample.
- Simulated grid (one setup only): one side at a time; an order is added at each one-minute open at least 0.01 USD worse than the last, with lots growing about 1.1 times; the basket closes at average price + 1 USD and before weekends; spread fixed at 0.10 USD; fills at bar opens, no margin limit. "Deep" means a virtual floating loss of 30,000 USD at a 0.1-lot base. 156,502 baskets in the main sample, 116,208 on the second feed.
- Tests: Jarque–Bera, Poisson, Hill tail index, binomial, Ljung–Box, Fisher, a yearly log-log slope, rate ratios and Spearman, with Holm and Benjamini–Hochberg corrections across all 11. Because extreme days cluster, we also resample whole calendar months (2,000 draws) and report those intervals.
3. Results
| Test | 2004–2025 | Second feed 2022–2026 |
|---|---|---|
| Days beyond 4 sd vs a normal law | 20 vs 0.34 expected, about 59× (41–89×) | 4 days, about 56× (14–99×) |
| Volatility clusters (lag-1 autocorrelation of squared returns) | 0.105, Holm p 5e-110 | 0.211 |
| Another extreme day within 5 days | 20.4% vs 4.4% (+16 points; +3 to +22) | 33.3% vs 3.8% (−4 to +47) |
| Extreme days that start from a quiet market | 13% (5–23%) | 25% (0–64%) |
| Tail index of daily returns (top 5%) | 3.42 (3.14–3.98) | 3.24 (2.60–4.69) |
| Tail index of basket adverse moves (top 1%) | 2.90 (2.79–3.05) | 2.39 (2.23–2.60) |
| Deep baskets, ATR20 ≥ 20 USD vs below | ratio 1.58 (1.36–1.82) | 1.08 (0.68–1.61) |
| Deep baskets, ATR20 as % of price, above vs below median | ratio 1.01 (0.88–1.17) | 1.12 (0.95–1.37) |
Ten of the 11 tests pass Holm at 5% in the main sample. The exceptions and weak points: the largest days are not reliably more often down (40 of 54, but not robust to the cut-off); the 1:1 growth of USD range with price is not rejected in 2004–2025 (slope 0.93); and the rank link between basket adverse moves and ATR is negligible in size (rho 0.03).
The price-level effect
Deep baskets look 58% more frequent when the 20-day ATR is 20 USD or more. Once ATR is measured as a share of price, the difference disappears in both feeds. The grid's depth limit and add step are fixed in dollars, so a basket goes deep more easily when gold is dearer, and dearer years are also years of high dollar ATR. The link is with the price level, not with volatility.
Range and the largest days
| Year | Range (USD) | Days beyond 3% |
|---|---|---|
| 2022 (from May) | 24.1 | 1 |
| 2023 | 23.7 | 2 |
| 2024 | 32.9 | 3 |
| 2025 | 60.7 | 5 |
| 2026 (to 9 Oct) | 122.8 | 20 |
The largest daily moves were +10.43% (17 September 2008) and −9.50% (15 April 2013) in the main sample, and −9.60% (30 January 2026) on the second feed. In 2025–2026 the dollar range rose faster than price (slope 1.74 on five yearly points; 1.45 and not significant without the partial 2026), so we report this as an observation, not a lasting change.
For the simulated grid, the worst adverse move from the first entry was 104 USD in 2004–2025 (9 November 2020) and 405 USD on the second feed (29 January 2026). These are virtual moves at bar opens; real fills when price gaps through several levels would differ.
4. What it shows
- Extreme days cluster and seldom start from quiet markets. The second feed is too short to confirm the last two points on its own.
- For this grid, adverse basket moves are heavy-tailed. We do not show that averaging causes this, or that other setups behave the same.
- For this grid, the apparent link between deep baskets and dollar volatility is explained by the price level. A setting fixed in dollars is relatively tighter when gold is dearer.
- Days beyond 4 sd are far more common than a normal law gives. This is about how often such days come, not the size of any portfolio loss.
5. Limitations
- Two feeds, each from one provider; the main feed has no spread and no stated time zone.
- The second feed overlaps the main one except for its last year, and is short (12 top-1% days).
- One simulated grid setup with fixed spread, bar-open fills and one-minute bars; not live trading.
- Month blocks correct for clustering only in part.
Sources
- Benjamini, Y., Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B 57(1), 289–300. https://doi.org/10.1111/j.2517-6161.1995.tb02031.x
- Hill, B. M. (1975). A simple general approach to inference about the tail of a distribution. Annals of Statistics 3(5), 1163–1174. https://doi.org/10.1214/aos/1176343247
- Holm, S. (1979). A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics 6(2), 65–70.
- novandraanugrah. XAU/USD Gold Price Historical Data. Kaggle dataset, CC0: Public Domain.
- Quantex Research WP 03: Gold calendar patterns: what survives a family-wide test and new data.
Declarations
- Author: Huyen X. Vu, Quantex Research, AlgoTrade Pte. Ltd., Singapore (ORCID 0000-0002-2209-842X).
- Type: Quantex internal research report, published on this site only; not peer reviewed. Reviewed before posting by two AI reviewers (Gemini, ChatGPT); all numbers re-checked against the script output.
- Use of AI tools: Data analysis and drafting were assisted by AI tools (Claude, Anthropic). The author checked the results and is responsible for the content.
- Conflict of interest: The author is the founder of AlgoTrade Pte. Ltd., which develops and runs systematic trading models.
- Funding: No external funding.
- Comments: welcome at [email protected].
- How to cite: Vu, H. X. (2026). Tail moves in gold, and what they mean for an averaging grid. Quantex Research, AlgoTrade Pte. Ltd. https://quantex.com.sg/research/gold-tail-moves/