Skip to content
#small language model Open access

Self-Normalization Brings In the Window You Divided By — The Layer Ratio Falls from 1.3795 Undivided to 1.0440 Divided, and the Median z Moves 4.3363 Times Across the Hours [F039]

Sep 2026 · Zenodo (CERN European Organization for Nuclear Research)

Abstract

「直近の平均で割って正規化した」。指標ではよく行われる。だが割り算は、分子と分母の両方を答えに入れる操作である。z が大きいのは分子が大きいからかもしれないし、分母が小さいからかもしれない。そして分母は「直近 N 本」から作られるので、z は N 本前までの相場の状態を必ず持ち込んでいる。本稿の分離子は「何で割ったか」である。分子・期間・銘柄・先の長さをすべて固定し、分母の窓の長さ N と、そもそも割るかどうかだけを変える。 四銘柄(USDJPY・EURUSD・GBPUSD・XAUUSD)の2019年から2024年までの1時間足で、分子 r = |直前の足の値動き|(pips)、分母 d = 直近 N 本の標準偏差(N = 16・64・256、直前の足までで作る)、z = r ÷ d とし、上位 5% を三通り(r で上位・d で下位・z で上位)取って、四時間先の値動きを比べた。まず、z の上位と r の上位は半分ほどしか重ならない。重なりは N=16 で 0.4968、N=64 で 0.6022、N=256 で 0.6695 である。そして向きが正反対である。z の上位では分母が全体の 0.6327 から 0.9330 倍と小さく、r の上位では 1.2989 から 1.4307 倍と大きい。z の上位の分子は全体の 3.7525 から 4.3332 倍で、r の上位の 4.8758 から 4.8803 倍より小さい。つまり割ると、分子で負けた分を分母の小ささで取り返した点、すなわち「静かな時期に起きた中くらいの動き」が上位に来る。 ここからが本稿の芯である。割らずに上位 5% を選ぶと四時間先の値動きは全体の 1.3795 倍になるのに、割ってから選ぶと 1.0440 倍、つまりほぼ全体と同じになる。割らない ÷ 割るは中央 1.3081 である。一方、分母だけで選ぶ(d の下位 5%)と 0.6635 倍と 1 を下回る。静かな時期のあとは先も静かだからである。割り算は、1 より上へ引く分子と 1 より下へ引く分母を混ぜており、その結果 z の層別はほぼ 1 に落ち着く。銘柄ごとに見ても同じで、USDJPY は r で 1.6855 倍・z で 1.0410 倍・d で 0.5387 倍、EURUSD の z は 0.9769 と 1 を下回る。N を伸ばすと z の層の比は 1.0440 → 1.0541 → 1.1145 と少し戻る。分母が動かなくなるほど、割らない場合に近づくからである。 さらに z は時間帯を持ち込む。N=64 での z の中央は、UTC 21 時の 0.3222 から UTC 14 時の 1.3972 まで 4.3363 倍動く。分母の窓が時間帯をまたぐので、静かな時間帯のあとに賑やかな時間帯が来ると分母がまだ静かなままで z が大きく出る。時間帯の最大 ÷ 最小は N=16 で 5.3367、N=64 で 4.5706、N=256 で 4.2731 で、窓を伸ばしても少し縮むだけで消えない。z の中央そのものも N で 0.6279 から 0.7927 まで動くので、「z が 1 を超えたら」という条件は N を決めなければ意味が定まらない。 そしてこれが全部、合成乱数では消える。割らない ÷ 割るは 0.9969(実測は 1.3081)、時間帯の最大 ÷ 最小は 1.1260(実測は 4.5706)、重なりは N=256 で 0.9029(実測は 0.6695)である。合成乱数では r で層別しても 0.9972 倍にしかならない。群れが無いので大きく動いた直後が特別ではなく、分母もほぼ一定なので割っても順位が変わらない。つまり本稿が測ったのは、ボラティリティの群れがある系列に特有の現象であって、割り算そのものの性質ではない。結論として、「正規化した」と書くときに必要なのは何で割ったかである。分母の窓の長さは前処理の細部ではなく、選ばれる時点そのものを決めている。 この論文の位置づけ:新しい数学定理も新しい法則も主張しない。正規化、ボラティリティの群れ、時間帯によるボラティリティの違いはすべて既知である。特定の教材やその著者、販売元は主題ではなく、一切名指していない。正規化を否定しない。方向を測っていない(扱うのは大きさだけである)。売買の成績もコストも測っていない。分子と分母を一つずつに固定している。時間帯の効果の機序を説明していない。上位を 5% に固定している。測定は日中・少数銘柄・四銘柄六年・1 時間足に限られ、月次の時系列モメンタムのようにより長い保有で文書化されている結果を否定するものではない。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、集計に用いたスクリプトの作成には AI(大規模言語モデル)の助力を得ました。最終的な内容の解釈や誤りがあれば、それらはすべて著者の責に帰します。お気づきの点があれば、ご教示いただければ幸いです。 ----- "Normalized by dividing by a recent average." Indicators do this all the time. But division puts both the numerator and the denominator into the answer: a large z may mean a large numerator, or it may mean a small denominator. And since the denominator is built from "the last N bars", z always brings in the state of the market up to N bars back. The separator in this paper is what you divided by. The numerator, the period, the instruments and the forward horizon are all held fixed, and only the window length N of the denominator — or whether to divide at all — is changed. Using one-hour bars for four instruments (USDJPY, EURUSD, GBPUSD, XAUUSD) from 2019 to 2024, with numerator r = |move of the previous bar| (pips), denominator d = the standard deviation of the last N moves (N = 16, 64, 256, built up to the previous bar) and z = r ÷ d, the top 5% was taken three ways (top by r, bottom by d, top by z) and the four-hour forward move compared. First, the top by z and the top by r overlap only about half: 0.4968 at N=16, 0.6022 at N=64 and 0.6695 at N=256. And the direction is exactly opposite. In the z top the denominator is 0.6327 to 0.9330 times the whole — small — while in the r top it is 1.2989 to 1.4307 — large. The numerator of the z top is 3.7525 to 4.3332 times the whole, below the r top's 4.8758 to 4.8803. So dividing brings in points that lost on the numerator and got in on the smallness of the denominator: a medium-sized move during a quiet stretch. Here is the core. Selecting the top 5% without dividing gives a four-hour forward move 1.3795 times the whole; selecting after dividing gives 1.0440 — that is, the same as the whole. Undivided ÷ divided has a median of 1.3081. Selecting by the denominator alone (the bottom 5% by d) gives 0.6635, below 1, because what follows a quiet stretch is quiet too. Division mixes a numerator that pulls above 1 with a denominator that pulls below it, and the z layer settles near 1. The same holds instrument by instrument: USDJPY gives 1.6855 by r, 1.0410 by z and 0.5387 by d, and EURUSD's z layer is 0.9769, below 1. Lengthening N brings the z layer ratio partly back, 1.0440 → 1.0541 → 1.1145, because the less the denominator moves the closer it gets to not dividing. Further, z brings in the hour. At N=64 the median z runs from 0.3222 at 21:00 UTC to 1.3972 at 14:00 UTC, a factor of 4.3363: the denominator's window straddles hours, so when a busy stretch follows a quiet one the denominator is still quiet and z comes out large. The hourly max ÷ min is 5.3367 at N=16, 4.5706 at N=64 and 4.2731 at N=256 — a longer window shrinks it only a little. The median z itself also runs from 0.6279 to 0.7927 depending on N, so a condition like "once z exceeds 1" has no meaning until N is named. And all of this disappears on a synthetic random walk. Undivided ÷ divided is 0.9969 (against 1.3081 on the measurement), the hourly max ÷ min is 1.1260 (against 4.5706), and the overlap at N=256 is 0.9029 (against 0.6695). On the random walk even the r layer gives only 0.9972: with no clustering the moment after a large move is not special, and the denominator is nearly constant, so dividing barely changes the ranking. What this paper measures is therefore specific to a series with volatility clustering, not a property of division itself. The conclusion: what is needed when writing "normalized" is what you divided by. The window length of the denominator is not a detail of preprocessing; it decides which moments get selected. Scope of this paper: no new mathematical theorem and no new law is claimed; normalization, volatility clustering and the intraday pattern of volatility are all known. No real course, author, or vendor is the subject here, and none is named. Normalization is not rejected. Direction is not measured — only magnitude. Neither trading performance nor cost is measured. One numerator and one denominator are fixed. No mechanism for the hourly effect is offered. The top share is fixed at 5%. The measurement is intraday and covers four instruments over six years on one-hour bars, and does not contradict results documented at longer holding periods, such as monthly time-series momentum. Acknowledgement: The ideas and content of this paper stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and writing the scripts used for the measurement. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated.

View source

Similar papers

#small language model Dataset Open access Oct 2026

Socratic guiding questions in synthetic arithmetic data: matched LoRA runs (revision v2)

Supporting data, adapters, predictions and code for the article *Low-Cost LoRA Fine-Tuning of Small Language Models for Multi-Step Arithmetic Reasoning* by Jake O'Grady, Asena Isik Gürhan, Chee Fong Ting and Effirul Ramlan (University of Galway). We generated 20,000 GSM8K-derived arithmetic problems with step-by-step s...

O'Grady, Jake, Gürhan, Asena Isik, Chee, Fong Ting et al. · 465 citations
#computer vision Open access Jun 2016

Software Development in Startup Companies: The Greenfield Startup Model

The results are packaged in the Greenfield Startup Model (GSM), which explains the priority of startups to release the product as quickly as possible, and the need to shorten time-to-market, by speeding up the development through low-precision engineering activities.

Carmine Giardino, Nicolò Paternoster, M. Unterkalmsteiner et al. · 178 citations · ⚡14
#computer vision Open access Oct 2016

Software Startups - A Research Agenda

Software startup companies develop innovative, software-intensive products within limited timeframes and with few resources, searching for sustainable and scalable business models.

M. Unterkalmsteiner, P. Abrahamsson, Xiaofeng Wang et al. · 157 citations · ⚡17
#machine learning Review Open access Oct 2016

“Failures” to be celebrated: an analysis of major pivots of software startups

This study conducts a case survey study based on the secondary data of the major pivots happened in 49 software startups, and demonstrates that customer need pivot is the most common among all pivot types.

Sohaib Shahid Bajwa, Xiaofeng Wang, Anh Nguyen-Duc et al. · 127 citations · ⚡15
#computer vision Review Open access May 2015

A survey study on major technical barriers affecting the decision to adopt cloud services

The comparison of adopter and non-adopter sample reveals three potential adoption inhibitor, security, data privacy, and portability, which underlines the importance of the technical and security perspectives for research investigating the adoption of technology.

Nattakarn Phaphoom, Xiaofeng Wang, S. Samuel et al. · 111 citations · ⚡8
#computer vision Open access Feb 2018

Lean Internal Startups for Software Product Innovation in Large Companies: Enablers and Inhibitors

This study investigates how Lean internal startup facilitates software product innovation in large companies and identifies its enablers and inhibitors, and shows the potential of the method-in-action framework to investigate the Lean startup approach in non-startup context.

Henry Edison, Nina M. Smørsgård, Xiaofeng Wang et al. · 78 citations · ⚡6

Related blog posts

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.