Skip to content
#small language model Open access

The Answer You Get After Removing the Overlap Depends on the Phase of the Thinning — Changing Only Where You Start Counting Splits the Sign in 0.8000 of the Cells, and 1.0000 at Horizons of 24 and 72 Bars [F024]

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

Abstract

候補を潰すとき、「上位5日を除いたら符号が反転した。だからこの候補は少数の日に乗っている」と書くことがある。本稿が点検するのは、著者自身がこれまで使ってきたこの検定である。上位を除けば、どんな列でも平均は必ず下がる。最大の観測を取り除くのだから、これは算術であって候補の脆さではない。したがって比べる相手が要る。自然な相手は、同じ日数を無作為に選んで除いた版である。無作為に除くのは不偏なので、平均は上がりも下がりもしないはずである。分離子はどの日を除いたかであり、日数もデータも期間も銘柄も固定したまま、選び方だけを変える。 四銘柄(USDJPY・EURUSD・GBPUSD・XAUUSD)の2019年から2024年までの1時間足で、四つの族(移動平均のクロス・RSI・バンドの外側・高安のブレイク)を先の長さ6・24・72本で掃引し、日次平均の平均が正のセルだけを見た。銘柄あたり 124 から 154 セル、日数の中央は 1,822 から 1,868 である。上位5日の寄与率は中央 0.6394 で、5 日は全体の 0.2682% にすぎない。上位10%日の寄与率は中央 7.3912 と 1 を大きく超えるが、これは合計そのものが小さいからで、寄与率は合計が 0 に近づくと発散する量である。「上位の数件が利益の九割」という言い方は、裾の重さを言っているようで合計の小ささを言っている場合がある。 本稿の芯は対照にある。上位5日を除くと 0.3076 のセルが 0 以下になり、上位10%日なら 0.8994 になる。ところが同じ日数を無作為に除くと、0.0000 と 0.0036 である。値そのものも 1.2847 が 1.2900 になるだけで動かない。落ちたのは、除き方の算術である。そして足の中まで独立な正規増分で作った合成系列、つまり方向の情報が定義からゼロの系列でも、上位5日を除けば 0.1376 のセルが 0 以下になり、寄与率の中央は 0.2853 ある。この検定は、当たっていない系列にも同じように効く。 寄与率が何で決まるかも測った。測る前に書いた予言のうち「日数が少ないほど寄与率が上がる」は外れ、相関は +0.0649 から +0.1175 の弱い正で、日数の多い半分の方が寄与率が高かった。そのかわり寄与率と「ばらつき ÷ 平均」の相関は 0.9990・0.9905・0.9995・0.9983 である。ただしこれは発見ではない。日数がほぼ一定なので、寄与率はその比のほぼ一次関数になる。つまり 0.99 はほぼ恒等式である。読み替えると、寄与率は「集中しているか」ではなく「一日あたりのばらつきが平均の何倍か」を測っている。相場ではその比がもともと 25.1954 から 43.2779 あるので、上位の数日が目立つのは既定である。実測の方が乱数より落ちる度合いは大きい(0 以下になる割合の差は中央 +0.1799)が、それは裾が重いからではない。ばらつきの中央は実測 55.0426 に対し乱数 59.6483 で乱数の方が大きく、違うのは平均の水準(1.7303 対 3.0074)である。しかもその平均の差自体が、正のセルだけを見たことの産物でありうる。本稿はその交絡を分離していない。結論は、この検定を捨てよということではない。少数の日に乗っているかを見る目的には合っている。言うのは、報告するときに、無作為に同じ日数を除いた対照と、全部で何日あったかを並べなければ読めない、である。 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。順序統計量、選択バイアス、ランダムウォーク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。売買の成績(手数料・約定・資金管理を含む)は測っていない。掃引の範囲を四つの族に固定しており、別の族や別の粒度では数字が変わる。正のセルだけを見ているので、実測と乱数の平均の差は選び方の産物でありうる。重複を除いていないので t 値を報告しておらず、有意性は論じていない。除き方を四通りに限っている。合成乱数はボラティリティの群れを含まず、群れがあれば乱数側の寄与率は上がって差は縮む方向に動くが、本稿はそれを測っていない。測定は日中・少数銘柄・四銘柄六年に限られ、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。 ----- When killing a candidate, one writes things like "dropping the top 5 days flipped the sign, so this candidate rides on a handful of days". What this paper inspects is that test, which the author himself has been using. Dropping the top must lower the mean of any series whatsoever — the largest observations are being removed — so this is arithmetic, not the fragility of the candidate. A comparison partner is therefore needed, and the natural one is the same number of days chosen at random: random dropping is unbiased, so the mean should neither rise nor fall. The separator is which days were dropped: the number of days, the data, the period and the instruments are held fixed, and only the choosing is varied. Using one-hour bars for four series (USDJPY, EURUSD, GBPUSD, XAUUSD) from 2019 to 2024, four families (moving-average cross, RSI, outside a band, high-low breakout) were swept at horizons of 6, 24 and 72 bars, keeping only cells whose mean of daily means is positive: 124 to 154 cells per instrument, with a median of 1,822 to 1,868 days. The top-5-day share has a median of 0.6394, while 5 days is only 0.2682% of the whole. The top-10%-day share has a median of 7.3912, far above 1, because the total itself is small: the share is a quantity that diverges as the total approaches zero. "The top few made ninety per cent of the profit" looks like a statement about the weight of the tail while sometimes being a statement about the smallness of the total. The core of the paper is the control. Dropping the top 5 days leaves 0.3076 of the cells at or below zero, and the top 10% leaves 0.8994. But dropping the same number at random leaves 0.0000 and 0.0036, and the value itself only goes from 1.2847 to 1.2900. What fell is the arithmetic of the dropping. And on a synthetic series built from independent normal increments down to the movement inside each bar — a series with zero directional information by construction — dropping the top 5 days still leaves 0.1376 of the cells at or below zero with a median share of 0.2853. The test works just as well on a series that is right about nothing. What the share is determined by was measured too. Of the predictions written in advance, "the fewer the days the larger the share" failed: the correlation is a weak positive, +0.0649 to +0.1175, and the half with more days had the higher share. Instead the share correlates with "spread ÷ mean" at 0.9990, 0.9905, 0.9995 and 0.9983. That is not a discovery, however: with the number of days nearly constant the share is very nearly a linear function of that ratio, so 0.99 is very nearly an identity. Read another way, the share measures not "is it concentrated" but "how many times the mean is the day-to-day spread", and in markets that ratio already runs from 25.1954 to 43.2779, so the top few days standing out is the default. The measurement does fall further than the random walk (a median difference of +0.1799 in the share left at or below zero), but not because its tail is heavier: the median spread is 55.0426 on the measurement against 59.6483 on the random walk — larger on the random walk — and what differs is the level of the mean, 1.7303 against 3.0074. That difference in mean may itself be a product of looking only at positive cells, a confound this paper does not separate. The conclusion is not to discard the test, which is fit for seeing whether a result rides on a handful of days; it is that a report of it cannot be read without the control that drops the same number at random, and the number of days there were in all, beside it. Scope of this paper: no new mathematical theorem and no new law is claimed; order statistics, selection bias and random walks are all known. No real course, author, or vendor is the subject here, and none is named. Trading performance, including commissions, fills and money management, is not measured. One sweep is fixed at four families, and other families or granularities give other figures. Only positive cells are looked at, so the difference in mean between the measurement and the random walk may be a product of that choice. The overlap is not removed, so no t value is reported and significance is not discussed. Only four ways of dropping are used. The synthetic series has no volatility clustering; with clustering the random walk's share should rise and the gap should narrow, but that was not measured. The measurement is intraday and covers four instruments over six years, 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.