masterofnothi.ng https://masterofnothi.ng/ quarto-1.10.18 Thu, 26 Feb 2026 23:00:00 GMT Traders trade information or they don’t https://masterofnothi.ng/posts/traders-trade-information/

Introduction

Online social trading platforms have created an ecosystem where individual traders publicly share their investment strategies and portfolio positions. Some participants operate as fund managers, accepting retail investor capital in exchange for management and performance fees. Others build personal brands, monetizing through subscriptions and affiliate marketing. Still others maintain public trading profiles without direct monetization. What these arrangements have in common is that there’s an online public record of trades.

From a standard economic perspective, these retail-focused strategies should underperform market indices. The Efficient Market Hypothesis [1] suggests that publicly available information is already incorporated into prices, leaving no exploitable inefficiencies for retail traders. Investors in these strategies would be expected to receive market-level returns with higher volatility and reduced net returns after fees — hardly an attractive proposition. And yet, when you look at these platforms, certain traders do seem to outperform the market consistently.

Figure 1: A social trading platform’s cumulative returns vs S&P 500 and EuroStoxx 50 from December 2, 2024 to January 29, 2026.
Annualized Volatility
Trading Platform 53%
S&P 500 18%
EuroStoxx 50 16%
Table 1: A social trading platform’s annualized volatility vs S&P 500 and EuroStoxx 50 from December 2, 2024 to January 29, 2026.
Correlation
S&P 500 17%
EuroStoxx 50 10%
Table 2: A social trading platform’s daily returns correlation with S&P 500 and EuroStoxx 50 from December 2, 2024 to January 29, 2026.

The obvious explanation is survivorship bias: successful traders remain visible while unsuccessful ones drop out from the platform. However, examination of user identification numbers reveals no systematic account deletion — IDs are assigned sequentially without large gaps that would indicate systematic removal of unsuccessful accounts. Additionally, many outperforming traders select low-attention, thinly-traded stocks that don’t fit typical retail preferences, which undermines explanations based purely on luck in picking popular growth stocks like Apple or Google.

This raises an interesting question: could these platforms facilitate the monetization of privileged information? We hypothesize that they may enable an information trading economy [2], where:

  • Traders receive privileged information from informants rather than possessing direct access themselves.
  • Informants share information with multiple traders rather than single recipients.
  • Traders further disseminate information among their networks, since information can be duplicated at zero marginal cost.

This study explores this hypothesis through quantitative analysis of trading network data. We examine whether clustering patterns and correlated trading activity among platform participants can be systematically identified, and whether such patterns predict subsequent price movements.

This study is conducted solely for personal purposes; it does not identify specific platforms or traders and makes no allegations of illegal activity against any individuals or entities.

Methods

Data

Our dataset consists of 2.2 million stock trades executed by 11,295 traders on an online social trading platform from January 1, 2023 to December 31, 2024. Daily closing prices for all ISINs were independently sourced from Yahoo Finance, where available. For each trade, the dataset contains the timestamp of the trade, an anonymized trader ID, the ISIN of the underlying stock, the stock price at which the trade was executed and the direction of the trade - buy or sell.

The dataset does not contain the size of the trades (i.e., the number of shares bought or sold) or any information linking specific buy and sell transactions. While we observe individual buy and sell trades with their timestamps and prices, we cannot determine which prior buy trade a trader is closing when they execute a sell trade. The platform does not allow short selling; therefore, all sell trades represent the closing of previous long positions, though the specific pairing cannot be identified from the available data. For our analysis, we aggregate trades at the monthly level by trader-stock pairs, calculating net positions to determine directional exposure rather than tracking individual trade-level returns.

Coordination Detection Methodology

We identify potential insider trading networks by analyzing coordinated trading patterns and their subsequent profitability. Our methodology consists of the following steps:

1. Monthly Trade Aggregation: For each calendar month from January 2023 to December 2024, we aggregate all trades by trader-stock pairs, calculating net positions (buy trades minus sell trades) to determine trading direction (long or short). We exclude neutral positions where net trades equal zero. For each trader-stock-month combination, we record:

  • Trade count: the number of trades executed.
  • Direction: long (net positive position) or short (net negative position).
  • Total trading activity: the trader’s total number of trades and unique stocks traded that month.

For each stock during each month, we calculate:

  • Stock popularity: the number of distinct traders who traded the stock.
  • Stock volume: the total number of trades executed on the stock.
  • Forward return: the percentage change in stock price over the subsequent month.

2. Coordination Identification: A trader pair is classified as coordinated when both traders take the same directional position (both long or both short) on the same stock during the same month. We generate all possible trader pairs who traded the same stock in the same month, creating a comprehensive record of potential coordination instances.

3. Profitability Classification: To assess whether coordinated trades preceded price movements, we classify each coordination instance as profitable or unprofitable. A coordinated trade is classified as profitable when the one-month forward return aligns with the trade direction: positive returns for long positions or negative returns for short positions. This forward-looking profitability metric captures whether the coordinated trading behavior predicted subsequent price movements.

4. Filtering for Suspicious Patterns: To identify the most suspicious coordination patterns, we apply multiple filters designed to isolate synchronization on illiquid stocks with abnormally high returns. We retain only trader pairs meeting all of the following criteria:

  • Stock illiquidity: The stock was traded by fewer than 100 unique traders during the analysis period.
  • Low trading volume: The stock had total trading volume below 1,000 trades during the analysis period.
  • Low trader activity: Each trader in the pair executed fewer than 1,000 total trades during the analysis period, filtering out hyperactive traders for whom coordination is more likely coincidental.
  • High profitability: The average profitable return when coordinating exceeded 20%, substantially above typical market returns.

These filters isolate trader pairs who repeatedly coordinate on multiple rare, illiquid stocks and achieve exceptional returns — a pattern consistent with advance knowledge of material non-public information rather than coincidental overlap or public information trading.

These filters are intentionally conservative and designed to isolate only the most extreme coordination patterns. If such patterns were purely coincidental, we would expect few or no persistent networks to emerge across independent time periods. The fact that similar network structures appear in both years suggests that the results are not driven solely by threshold choice.

Network Construction and Analysis

We construct an undirected network graph where nodes represent traders and edges represent coordination relationships. An edge connects two traders if they coordinated on at least one stock meeting the filtering criteria above. Each edge is weighted by the average profitable return achieved across all stocks on which the trader pair coordinated. This network structure allows us to identify clusters of traders who consistently coordinate with each other on profitable trades, potentially indicating information-sharing networks.

For each trader in the network, we calculate:

  • Degree: The number of other traders with whom the trader coordinated.
  • Degree centrality: The proportion of all possible connections the trader has, normalized by network size.
  • Node return: The average profitable return across all coordination edges connected to the trader.

We also identify connected components — isolated subgraphs of traders who are connected to each other but disconnected from the rest of the network. The largest connected component represents the primary cluster of coordinated traders.

To visualize the relationship between traders and stocks, we construct a bipartite network where trader nodes and stock nodes are connected by edges representing trading activity. This representation reveals which stocks serve as common coordination points across multiple traders, highlighting potentially suspicious securities that may be the subject of insider information.

Analysis Period

We conduct separate analyses for calendar years 2023 and 2024 to identify temporal patterns in coordinated trading behavior and assess whether suspicious networks persist across different time periods.

Results

2023 Analysis

For the period from January 1, 2023 to December 31, 2023, our filtering criteria identified a network of synchronized traders operating on illiquid stocks with abnormally high returns. The resulting trader network exhibits specific structural properties that suggest information sharing rather than coincidental simultaneous positioning.

Network Structure and Centrality

Figure 2 presents the relationship between network position and trading performance. The top panel shows average profitable returns as a function of degree (the number of coordination partners each trader has). We observe a weak positive relationship (β = 0.0111, p = 0.4148), though not statistically significant. The bottom panel examines the relationship with degree centrality (the proportion of possible connections each trader maintains). Here we find again a positive relationship (β = 1.3305, p = 0.4148), still not reaching statistical significance.

Figure 2: Network metrics from January 1, 2023 to December 31, 2023.

While these relationships do not achieve statistical significance, the positive coefficients suggest a pattern where more connected traders tend to achieve higher returns, consistently with the findings in the literature.

Most Profitable Coordination

Figure 3 illustrates the most profitable coordinated trade identified in 2023: two traders executing synchronized long positions in Fusion Fuel Green PLC (ISIN: IE00BNC17X36) during November 2023. The green marker indicates the timing of coordinated entry, positioned ahead of a dramatic price surge from approximately $25 to over $55 in December 2023 — a return exceeding 100% within one month. This trading pattern is difficult to reconcile with standard momentum-based or technical trading strategies. Fusion Fuel Green PLC is a thinly-traded renewable energy company with limited analyst coverage and minimal retail investor attention, making simultaneous independent discovery by multiple traders unlikely.

Figure 3: Most profitable trade executed between January 1, 2023 and December 31, 2023.

Largest Coordination Network

Figure 4 shows the stock with the most coordinated traders in 2023, Semler Scientific, Inc. (ISIN: US81684M1045), where six traders took long positions in November 2023. The synchronized entry (indicated by the green marker) preceded a steady upward trend from approximately €29 to €46 by late December 2023, representing a ~59% return. While less dramatic than the Fusion Fuel example, the aligned behavior of six independent traders on this small-cap healthcare technology and diagnostics company — again, not a typical retail favorite — ahead of a sustained price increase could indicate advance knowledge of positive developments.

Figure 4: Trade with the most coordinated traders between January 1, 2023 and December 31, 2023.

Network Topology

Figure 5 presents the bipartite network structure of the largest connected component in 2023. Blue circles represent traders, while orange squares represent stocks. This visualization reveals two stocks — Twist Bioscience Corporation and CytoMx Therapeutics, Inc. — connecting multiple traders. Both companies operate in the biotechnology sector, suggesting potential information flow from sources with access to biotech industry developments.

Figure 5: Largest connected component between January 1, 2023 and December 31, 2023.

2024 Analysis

Analysis of the 2024 trading period reveals similar patterns, with some notable differences in network structure and the stocks serving as coordination points.

Network Structure and Centrality

Figure 6 presents network metrics for 2024. The relationship between degree and average returns (top panel) shows a positive but not statistically significant relationship (β = 0.0068, p = 0.7822). The relationship with degree centrality (bottom panel) is similarly positive but not significant (β = 1.2122, p = 0.7822). Notably, the 2024 network exhibits higher baseline returns (30-50% compared to 30-35% in 2023) but with wider confidence intervals, suggesting greater heterogeneity in coordination profitability.

Figure 6: Network metrics from January 1, 2024 to December 31, 2024.

Most Profitable Coordination

Figure 7 shows the most profitable 2024 coordination: five traders entering long positions in Vuzix Corporation (ISIN: US92921W3007) in December 2024. The entry preceded a dramatic price surge from $2.50 to over $5.50 — more than doubling in value within weeks. Vuzix, a manufacturer of augmented reality devices, represents another example of synchronized trading on an obscure technology company with limited mainstream attention.

Figure 7: Most profitable trade executed between January 1, 2024 and December 31, 2024.

Largest Coordination Network

Figure 8 presents the stock with the most coordinated traders in 2024: Reconnaissance Energy Africa Ltd (ISIN: CA75624R1082), involving seven traders who entered long positions in May 2024. The entry occurred at a price of around CA$0.90, followed by a sustained rally to CA$1.45 by June 2024. Reconnaissance Energy Africa, an oil and gas exploration company operating in Namibia, represents yet another thinly-traded stock in a specialized sector unlikely to attract coincidental simultaneous retail trader attention.

Figure 8: Trade with the most coordinated traders between January 1, 2024 and December 31, 2024.

Network Topology

Figure 9 shows the largest connected component in 2024, again structured around two coordination hub stocks: Reconnaissance Energy Africa Ltd and Chesapeake Gold Corp. The network topology is similar to 2023 — multiple traders connected through common trading of two stocks in related sectors (in this case, natural resources exploration). Interestingly, the sector focus shifts from biotechnology (2023) to natural resources exploration (2024), suggesting that the information sources feeding these networks may vary across time, or that different information networks operate in different sectors.

Figure 9: Largest connected component between January 1, 2024 and December 31, 2024.

Discussion

The consistency of the results across 2023 and 2024 — despite partially different stocks and traders and varying market conditions — suggests that we are observing a pattern of information sharing rather than coincidental synchronicity. The positive relationship between network centrality and returns, while not statistically significant, appears in both years with consistent directionality, suggesting limited statistical power rather than absence of effect.

Alternative explanations merit consideration:

Influencer following: Traders might be following influential platform personalities. However, popular influencers typically discuss well-known stocks rather than obscure small-caps, and would produce larger, more visible coordination networks than we observe.

Shared analytical approach: Independent analysis using similar frameworks might produce synchronization. However, this cannot explain the precise timing we observe — multiple traders entering simultaneously at price inflection points on thinly-traded stocks with limited public information.

Platform algorithms: Recommendation systems might suggest similar stocks to similar users. However, platforms typically recommend high-volume stocks, not the illiquid securities we identify, and algorithmic recommendations would produce more diffuse timing patterns.

More granular temporal analysis using daily or hourly trade data could reveal whether coordination occurs simultaneously or sequentially, informing whether traders receive information from common sources or from each other. Cross-platform analysis could determine whether networks span platforms or remain platform-specific. Linking coordinated trades to subsequent news announcements or corporate events would provide more direct evidence of information advantage.

Appendix

The network analysis showed that predicting specific instances of insider trading is not possible, as coordination events are extremely rare and typically occur only once per trader pair, each time involving different stocks, leaving no historical pattern of repeated coordination from which to predict which traders will coordinate again or which stocks they will target next.

We can nevertheless leverage the collective signal from subsets of high-performing traders to predict future price movements of individual stocks. In the rest of this section, we develop and validate a machine learning model to predict the 30-day-ahead return of individual stocks using a set of indicators that capture changes in positions and trade volumes across a set of selected traders. The historical correlation between the change in a trader’s position and the 30-day-ahead return is used to identify the traders whose activity is most predictive of future price movements.

A trader’s position on a given day is defined as the cumulative number of buy trades minus the cumulative number of sell trades prior to and including that day. The change in a trader’s position is defined as the average position over the previous 30 days minus the average position over the previous 90 days: this increases when buy trades outnumber sell trades - indicating that the trader is taking a long position on the stock - and decreases when sell trades outnumber buy trades - indicating that the trader is taking a short position on the stock.

The following indicators are used as features to predict the 30-day-ahead return:

  • Buy Trades Change: Change in the 30-day average of daily total buy trades compared to the 90-day average.
  • Sell Trades Change: Change in the 30-day average of daily total sell trades compared to the 90-day average.
  • Long Traders Change: Change in the 30-day average of daily counts of traders with more buy than sell trades compared to the 90-day average.
  • Short Traders Change: Change in the 30-day average of daily counts of traders with more sell than buy trades compared to the 90-day average.
  • Largest Buy Trade Change: Change in the 30-day average of the largest number of buy trades executed on a single day compared to the 90-day average.
  • Largest Sell Trade Change: Change in the 30-day average of the largest number of sell trades executed on a single day compared to the 90-day average.
  • Largest Buy Weight Change: Change in the 30-day average of the largest portfolio weight for a buy trade compared to the 90-day average.
  • Largest Sell Weight Change: Change in the 30-day average of the largest portfolio weight for a sell trade compared to the 90-day average.
  • Best Performance Change: Change in the 30-day average of the best trade performance compared to the 90-day average.
  • Worst Performance Change: Change in the 30-day average of the worst trade performance compared to the 90-day average.

We use the above features to train a CatBoost classifier which predicts the sign of the 30-day-ahead return (class 1: positive return, class 0: negative return) of a given stock. The hyperparameters used for training the CatBoost algorithm are reported in Table A1.

Hyperparameter Value
objective Logloss
iterations 100
boosting_type Ordered
grow_policy SymmetricTree
has_time True
l2_leaf_reg 3
subsample 0.7
rsm 0.7
Table A1: CatBoost hyperparameters.

We backtest the model by training it on each day from January 1, 2025 to September 30, 2025 using the most recent 5 years (5 x 250 days) of data, and using the trained model for predicting the probability that the stock price will go up over the subsequent 30 days. For each stock, we extract the features using only the activity of the traders whose position changes have at least 10% correlations with the 30-day-ahead stock return over the 5-year training set.

The results indicate that the model has good predictive ability on a set of highly traded ISINs, as both the accuracy score and ROC-AUC score are above 60% (Figure A1, Figure A2, Figure A3 and Figure A4). Further research is needed to assess whether these results generalize across a broader set of stocks and over different time windows.

Figure A1: Results for First Solar (ISIN: US3364331070).
Figure A2: Results for Palantir Technologies (ISIN: US69608A1088).
Figure A3: Results for Amazon (ISIN: US0231351067).
Figure A4: Results for Microsoft (ISIN: US5949181045).

References

[1]
E. F. Fama, “Efficient capital markets: A review of theory and empirical work,” The Journal of Finance, vol. 25, no. 2, pp. 383–417, 1970, doi: 10.2307/2325486.
[2]
D. Duffie, S. Malamud, and G. Manso, “Information percolation in segmented markets,” Journal of Economic Theory, vol. 153, pp. 1–32, 2014, doi: 10.1016/j.jet.2014.05.006.
[3]
K. R. Ahern, “Information networks: Evidence from illegal insider trading tips,” Journal of Financial Economics, vol. 125, no. 1, pp. 26–47, 2017, doi: 10.1016/j.jfineco.2017.03.009.
[4]
L. Adamic, C. Brunetti, J. H. Harris, and A. Kirilenko, “Trading networks,” The Econometrics Journal, vol. 20, no. 3, pp. S126–S149, 2017, doi: 10.1111/ectj.12090.
[5]
H. N. Ozsoylev, J. Walden, M. D. Yavuz, and R. Bildik, “Investor networks in the stock market,” The Review of Financial Studies, vol. 27, no. 5, pp. 1323–1366, 2014, doi: 10.1093/rfs/hht065.
]]> https://masterofnothi.ng/posts/traders-trade-information/ Thu, 26 Feb 2026 23:00:00 GMT How many werewolves are there, really? https://masterofnothi.ng/posts/werewolf-scan/

Introduction

Do you ever watch the news or read the papers after a busy day of doing nothing, then pause, stare at the ceiling, and think about geopolitics, the economy, human conflict - all of it - and wonder “what next”?

Everything happening right now is helping a third party. Who actually benefits from a tariff war? From long-running regional conflicts like Russia and Ukraine? From everyone obsessing over AI risk? Werewolves.

A lot of the news lately feels like one of those movie scenes where friends argue about who forgot to bring wine to the party. It’s obvious to the viewer this isn’t the actual plot - just a preamble. They’re about to get hit with something. And your brain goes: “werewolf attack?” The timing’s getting too good for them not to strike.

Figure 1: A werewolf.

As if further confirmation were needed, I recently looked at the silver price chart and saw exactly what I feared: a suspicious spike. Who’s buying all that silver? Potentially “bearish” investors - if you know what I mean, them having fur and everything. The recent price development looks suspiciously like late-stage attack planning. Everyone knows a werewolf can only be killed with a silver bullet. They’re buying up the silver so we can’t make bullets.

Figure 2: Silver price in US Dollars from January 1, 2019 to January 13, 2026 (Source: investing.com).

In this paper, I hypothesize that werewolves are using our internet and social networks. Since werewolves become aroused during a full moon, they’ll type more sloppily under its influence. To confirm this threat exists, I investigate the correlation between spelling and grammar errors in X.com (formerly Twitter) and the lunar cycle. The correlation’s strength will reveal their numbers - what we’re up against.

There’s ample evidence human spelling isn’t affected by the lunar cycle. Machines (bots, AI tools) aren’t affected either. This makes it a good test for detecting something using our networks that is neither human nor machine.

Methods

I performed the correlation analysis over two time windows: December 26, 2019 to February 23, 2020 (first scan) and December 14, 2025 to January 13, 2026 (second scan).

I scraped all posts containing the following hashtags:

  1. #beautiful
  2. #beer
  3. #birthday
  4. #challenge
  5. #competition
  6. #Dublin
  7. #food
  8. #foodie
  9. #happy
  10. #ireland
  11. #work

These hashtags were selected for their high posting frequency and geographic diversity, maximizing our chances of capturing werewolf activity across different time zones and social contexts.

For the first scan, I used TwitterScraper. For the second scan, I used Octoparse Twitter Scraper, as TwitterScraper is no longer maintained - possibly due to werewolf interference with open-source development, though this remains speculative.

I filtered for English-language posts and cleaned the text by removing URLs and emojis. I then analyzed each post for grammar and spelling errors using LanguageTool [1]. For the first scan, I used GAMET [2], which relies on Java LanguageTool. For the second scan, I used the Python wrapper for LanguageTool.

I grouped the data by day and calculated the total number of language errors and total number of posts for each day, both by hashtag and across all hashtags. I used this to calculate the daily error rate: total language errors divided by total posts.

For both scans, I obtained daily lunar data using the suncalc R package [3]. I calculated Pearson correlation coefficients between the daily error rate and four binary variables representing lunar phases: 1st Quarter, New Moon, 3rd Quarter, Full Moon, as well as three continuous variables: moon angle, moon fraction and moon phase.

Results

First Scan

The first scan revealed no statistically significant correlation between error rate and the lunar cycle. The absolute correlations calculated for each hashtag with the four binary variables - 1st Quarter, New Moon, 3rd Quarter, Full Moon - are not particularly large, and none reach statistical significance (Table 1). Similarly, the correlations with the three continuous variables - moon angle, moon fraction and moon phase - are relatively small and not statistically significant (Table 2).

While this fails to confirm the werewolf hypothesis, it does not exclude it. The absence of correlation could indicate sophisticated countermeasures - werewolves may be employing spellcheck or limiting social media activity during full moons. A larger sample or different time window might be required to detect their digital activity.

Hashtag Full Moon 3rd Quarter New Moon 1st Quarter
#beautiful 0.3767 0.2343 0.0622 0.1076
#beer 0.2082 0.1864 0.1994 0.1594
#birthday 0.3191 0.0884 0.0357 0.2203
#challenge 0.0101 0.1207 0.0555 0.0585
#competition 0.2406 0.4030 0.1455 0.0101
#Dublin 0.1486 0.1207 0.0200 0.3041
#food 0.0976 0.2543 0.1160 0.0415
#foodie 0.1085 0.1061 0.0448 0.0525
#happy 0.1761 0.1437 0.3183 0.2664
#ireland 0.0336 0.0039 0.0948 0.0518
#work 0.2140 0.1361 0.1486 0.0521
Table 1: Absolute correlations between daily error rates and binary lunar phase variables from December 26, 2019 to February 23, 2020.
Variable Correlation p-value
Moon angle 0.0745 0.5715
Moon fraction -0.1538 0.2408
Moon phase -0.1003 0.4459
Table 2: Correlations between daily error rates and continuous lunar variables from December 26, 2019 to February 23, 2020.
Figure 3: Error count from December 26, 2019 to February 23, 2020.
Figure 4: Error rate vs moon angle from December 26, 2019 to February 23, 2020.
Figure 5: Error rate vs moon fraction from December 26, 2019 to February 23, 2020.
Figure 6: Error rate vs moon phase from December 26, 2019 to February 23, 2020.

Second Scan

Five years later, I conducted a second scan. The first scan’s results were inconclusive, and the situation may have changed. The initial methodology may have been too obvious. I suspect werewolves saw it coming and deployed countermeasures - specifically, posting deliberately misspelled tweets with negative correlation to the lunar cycle to mask the real signal. The null results from the first scan suggest such counter-intelligence operations may exist.

A second scan makes sense for several reasons:

  1. Personnel turnover: The werewolf operative responsible for countermeasures may have quit, and recruitment for such expert-level appointment has likely proven challenging.
  2. Leave of absence: They may be on holiday or parental leave. Coordinating coverage for a position this specialized is nearly impossible.
  3. Professional burnout: After years of unrecognized work posting misspelled tweets into the void, junior operatives may have become demoralized and disengaged.
  4. Budget cuts: Their program may have lost funding. It’s hard to justify expenses when you can’t show results.
  5. Regulatory shutdown:. They may have been shut down for compliance reasons - perhaps failing a data protection audit.
  6. Technical complexity: Their countermeasure system must index misspelled tweets during full moons to werewolf-to-human X.com usage ratios, adjusted for changing usage intensity. This is complex and prone to calibration errors.

The second scan yielded results nearly identical to the first: no statistically significant correlations between error rates and lunar phases (Tables 3 and 4).

The pattern - or lack thereof - mirrors the first scan almost exactly. This consistency across a five-year gap cannot be coincidental. The countermeasures are still active. The fact that they maintained operational security for five years while we were dormant suggests sophisticated coordination. They knew we might return. This level of preparation indicates the threat is larger and better organized than initially hypothesized.

Hashtag Full Moon 3rd Quarter New Moon 1st Quarter
#beautiful 0.3166 0.0154 0.6102 0.3013
#beer 0.0781 0.1590 0.4289 0.3667
#birthday 0.2127 0.1412 0.3693 0.0397
#challenge 0.1044 0.0929 0.5053 0.3275
#competition 0.1624 0.1602 0.0339 0.0753
#Dublin 0.0062 0.2819 0.4778 0.1524
#food 0.1339 0.0990 0.3407 0.0876
#foodie 0.0018 0.0288 0.3744 0.3962
#happy 0.1320 0.0862 0.3057 0.3236
#ireland 0.0816 0.0666 0.3227 0.3773
#work 0.0366 0.0359 0.3138 0.3775
Table 3: Absolute correlations between daily error rates and binary lunar phase variables from December 14, 2025 to January 13, 2026.
Variable Correlation p-value
Moon angle 0.1955 0.2920
Moon fraction 0.1255 0.5013
Moon phase 0.2723 0.1384
Table 4: Correlations between daily error rates and continuous lunar variables from December 14, 2025 to January 13, 2026.
Figure 7: Error count from December 14, 2025 to January 13, 2026.
Figure 8: Error rate vs moon angle from December 14, 2025 to January 13, 2026.
Figure 9: Error rate vs moon fraction from December 14, 2025 to January 13, 2026.
Figure 10: Error rate vs moon phase from December 14, 2025 to January 13, 2026.

Discussion

The absence of statistically significant correlations in both temporal scans warrants consideration of alternative explanations.

The werewolf defense system may be more sophisticated than assumed, potentially employing browser-based spellcheck plugins that alert users to errors before posting, or implementing behavioral protocols that restrict internet usage during full moons entirely.

Alternatively, werewolves may not use X.com as their primary communication platform, instead maintaining separate social networks inaccessible to external observation - rendering our sampling methodology fundamentally ineffective.

The premise that werewolves are cognitively affected by lunar phases may itself be erroneous. They may simply possess enhanced night vision during full moons, providing a hunting advantage that has been misattributed to lunar sensitivity in the historical literature. Additionally, transformed werewolves may have more pressing activities during full moons than internet usage, making social media behavior an inappropriate detection proxy.

Several methodological concerns must be acknowledged. The analysis may contain unintentional errors in data processing or statistical calculation. More concerning is the possibility of intentional manipulation. If the author has been compromised - either through coercion or payment in silver - the integrity of these findings cannot be assumed. Related to this, if the author is themselves a werewolf, the entire analytical framework becomes suspect. Alternatively, the author may represent a fourth party - neither human, machine, nor werewolf - with strategic interests in fomenting interspecies conflict.

More sophisticated explanations involve infrastructure penetration. Werewolves may have infiltrated X.com’s systems, enabling real-time detection of scraping activities and redirection to curated data that has been stratified to eliminate temporal patterns in error rates. Even more concerning is the possibility of DNS infrastructure compromise, allowing werewolves to intercept and redirect data requests at the network level for any traffic originating from known research sources.

These competing explanations cannot be definitively resolved with current data. Further research is required, though studying an adversary with potential information superiority presents inherent challenges.

References

[1]
D. Naber, “A rule-based style and grammar checker,” Diplomarbeit, Technische Fakultät, Universität Bielefeld, 2003.
[2]
S. Crossley, A. Bustamante, and F. Bradfield, “GAMET: A tool for automatically assessing grammar and mechanic errors in learner corpora,” in 14th american association for corpus linguistics (AACL) conference, 2019, p. 16.
[3]
B. Thieurmel and A. Zeileis, suncalc: Compute sun position, sunlight phases, moon position and lunar phase. 2019.
]]> https://masterofnothi.ng/posts/werewolf-scan/ Mon, 12 Jan 2026 23:00:00 GMT