Skip to content
#small language model Open access

A Naive t Value Inflates with the Number of Observations per Day: Changing the Bar Size Moves the Count by a Factor of 11.9751, and Even a Synthetic Series with No Edge Exceeds 2 in 0.5437 of Five-Minute Cells [F014]

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

Abstract

手法の成績に t 値が添えられていることがある。その多くは、標準偏差を件数の平方根で割ったものを分母にした素朴な t である。だが同じ日の観測は独立でない。同じ相場の同じ動きを、何度も数えているからである。一日あたり m 件、日の中の相関を ρ とすると、標準誤差の比は √(1 + (m − 1)ρ) になる。ρ が 0 なら膨らまず、1 なら √m 倍になる。 四銘柄(USDJPY・EURUSD・GBPUSD・XAUUSD)の2022年から2024年までを、五分足・十五分足・一時間足の三つの足種で測った。先の長さは三つとも 24 時間にそろえてあるので、変わるのは一日あたりの件数だけである。信号は符号を持つ四つの族(直前の符号・移動平均のクロス・RSI の 50・バンドの外側)である。一日あたりの件数の中央は五分足 227.9935・十五分足 76.0392・一時間足 19.0390 で、11.9751 倍動く。日の中の相関の中央は 0.1430、幅は +0.0010 から +0.6348 でゼロではない。標準誤差の比の中央は 2.5265、幅は 1.1070 から 6.1256 である。ただし ρ をこの式から逆算している以上、比と式は恒等式であって検定ではない。本稿が測ったのは式の正しさではなく、ρ と m の実際の値である。なお比を上限 √m で割った値の中央は 0.4055 で、√m を目安に割り引くのは割り引きすぎである。 同じデータ・同じ信号で分母の作り方だけを変えると、素朴な t の絶対値が 2 以上になるセルは 48 のうち 19、日次クラスタでは 48 のうち 6 になる。中央は 1.8046 と 0.7218 である。差は足種が細かいほど大きく、五分足では 8/16 が 1/16 に落ちるのに、一時間足では 4/16 が 3/16 にしか落ちない。件数の多い手法ほど素朴な t は膨らむので、件数の多い手法と少ない手法の t を素朴なまま並べて比べてはいけない。USDJPY の十五分足・バンドの外側では、素朴な t が −3.2830 なのに日次 t が 0.0574 で、符号まで変わる。 本稿の芯は基準線にある。足の中まで独立な正規増分で作った合成系列は、方向の情報がゼロであることが構成から分かっている。そこで素朴な t の絶対値の中央は五分足で 2.0893、2 以上になった割合は 0.5437 である。半分以上のセルが「有意」に見える。したがって素朴な t は「当たっているか」を測っていない。測っているのは、同じ日を何回数えたかである。そして日次クラスタにしても足りない。同じ合成乱数で、一時間足の日次 t の絶対値の中央は 1.3530 で、五分足の 0.8516 より大きい。24 時間先まで見ているので、日をまたぐ重なりが残るためである。 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。級内相関、クラスタ頑健な標準誤差、重複した観測が t を膨らませること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。標準誤差の比が √(1 + (m − 1)ρ) になることは恒等式であり、本稿はそれを確かめたとは書かない。重複を除いていないので、本稿は「日次クラスタにすれば十分だ」とは言っていない。合成乱数での三つの足種が単調でなく、十五分足だけ値が小さい理由を説明できていない。信号の族を四つに固定しており、ρ を日を群とした一つの定義でしか測っていない。測定は日中・少数銘柄・四銘柄三年に限られ、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。 ----- A method's results often come with a t value, and most of them use the standard deviation divided by the square root of the count as the denominator. But observations inside one day are not independent: they count the same movement of the same market several times. With m observations a day and a within-day correlation, the ratio of standard errors is the square root of 1 + (m − 1) times that correlation — no inflation at zero, a factor of the square root of m at one. Four series (USDJPY, EURUSD, GBPUSD, XAUUSD) were measured from 2022 to 2024 on five-minute, fifteen-minute and hourly bars. The horizon is held at 24 hours across all three, so only the number of observations per day changes. Four signal families that carry a sign were used (the last bar's sign, a moving-average cross, RSI against 50, and outside a Bollinger band). The median observations per day are 227.9935, 76.0392 and 19.0390, a factor of 11.9751. The median within-day correlation is 0.1430, ranging from +0.0010 to +0.6348, and is not zero. The median ratio of standard errors is 2.5265, ranging from 1.1070 to 6.1256. Since the correlation is solved from that very relation, however, the ratio and the formula are an identity, not a test: what this paper measured is not the correctness of the formula but the actual values of the correlation and of m. The ratio divided by the upper bound √m has a median of 0.4055, so discounting by √m as a rule of thumb discounts too much. With the same data and the same signals, changing only the denominator, the naive t reaches 2 in absolute value in 19 of 48 cells and the daily-clustered t in 6 of 48, with medians of 1.8046 and 0.7218. The gap is larger on finer bars: on five-minute bars 8/16 falls to 1/16, while on hourly bars 4/16 falls only to 3/16. The more observations a method produces, the more the naive t inflates, so the t values of a high-count and a low-count method must not be compared while both are naive. For the band signal on USDJPY fifteen-minute bars, the naive t is −3.2830 while the daily t is 0.0574 — even the sign changes. The core of the paper is the baseline. A synthetic series built from independent normal increments down to the movement inside each bar has zero directional information by construction. There the median absolute naive t on five-minute bars is 2.0893 and the share exceeding 2 is 0.5437: more than half the cells look "significant". The naive t is therefore not measuring whether anything is right; it is measuring how many times the same day was counted. And daily clustering is not enough either: on the same synthetic series the median absolute daily t on hourly bars is 1.3530, larger than the 0.8516 of five-minute bars, because a 24-hour horizon leaves overlap that crosses days. Scope of this paper: no new mathematical theorem and no new law is claimed; intraclass correlation, cluster-robust standard errors, and the fact that overlapping observations inflate t 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. That the ratio of standard errors equals the square root of 1 + (m − 1) times the correlation is an identity, and this paper does not write that it confirmed it. The overlap is not removed, so the paper does not say that daily clustering is sufficient. The three bar sizes are not monotone on the synthetic series and the paper cannot explain why the fifteen-minute row is small. Four signal families are fixed and the correlation is measured under one definition, with days as groups. The measurement is intraday and covers four instruments over three 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.