81: Chapter 81 Beyond the Rankings

After finishing dinner and returning to his room, the computer screen had already entered sleep mode due to a long period of inactivity.

Jiang Lin shook his mouse, and the screen lit up; the progress bar in the download manager was already full.

That dataset, labeled minute_market_data_v3.zip with a size of a full 3.4 GB, lay quietly in his local hard drive.

Taking advantage of the gap while unzipping, Jiang Lin looked over the rules.

On the platform's task description page, which had an extremely crude UI, the rules were written very long, consisting of a dense, massive wall of disclaimer clauses and scoring details.

The data was prohibited from being leaked externally, public algorithms were allowed, and it was permitted to submit cleaned anomaly-marked files, method description reports, auxiliary scripts, and so on.

Jiang Lin copied several sections of the rules, pasted them into a newly created local notepad, and then used his mouse to bold three words among them.

Hidden labels, reproducibility, generalization capability.

These three words were much more important than that eight-thousand-yuan bonus.

They revealed the organizers' true pain points.

They didn't want those competition-exclusive models that scored very high on a specific dataset but completely flopped the moment the market environment changed.

What they wanted was an industrial-grade cleaning pipeline that could truly be implemented in practice.

Yet Jiang Lin furrowed his brows slightly as he read.

Because what he was familiar with was bad data, not the financial market.

Temperature sensor zero drift, camera dark-field noise, thermal expansion and contraction of optical mounts, and abnormal spikes knocked out of radiometers by cosmic rays—he had dealt with these things far too many times.

But minute-level market quotes were not temperature curves, trading volume was not sensor sampling, and trading suspension marks were not equipment dropouts.

If one hadn't even figured out the basic working manner of the system yet rushed to start cleaning data, that couldn't be called auditing.

That was called writing one's ignorance into scripts.

Jiang Lin temporarily put the download page aside and created a new document.

The title consisted of only four words: Quantitative Remedial Lessons.

Objective: Learn under what systems and processes these data were generated.

He started checking from the most basic fields first.

Opening price, trading volume, rights adjustment, price limits...

For every single word, he only looked at the most basic definitions without diverging into investment strategies.

When he encountered posts talking about tenfold returns, stable arbitrage, and perfect capital curves, he closed them directly.

When he encountered long posts where people seriously discussed future functions, survivorship bias, out-of-sample failure, slippage, and impact costs, he would stop to read them.

Two hours later, he listed six institutional issues that had to be confirmed first, ranging from call auction boundaries to test set isolation methods, and drew a very rough data pipeline on paper.

His final evaluation was: "Quantitative data is not the market itself; it is the shadow left behind after the market passes through multiple layers of equipment, institutions, software, and commercial calibers."

After finishing this sentence, he downloaded the data dictionary and the Baseline (benchmark reference code) provided by the platform.

The data dictionary was a PDF file of more than twenty pages.

He had to admit that the provider did an absolute job on desensitization.

That so-called target code was by no means any stock code like 600519 that could be searched within software, but was entirely converted into a long string of completely random MD5 hash values.

The trading days (date_id) in the timestamps were also shifted overall. You had no way of knowing whether this was data from the 2015 stock market crash or data from the 2020 herd-behavior market trend.

All prices and trading volumes had also undergone proportional scaling in the backend; their absolute values had lost meaning, retaining only zero values, missing values, adjacent rates of change, and relative relationships under the same caliber.

This meant that it was impossible for you to rely on the financial common sense of the Real World or look up historical records on Yahoo Finance to deduce backwards which stock experienced what anomaly on a certain month, day, and year.

This suited Jiang Lin's tastes very well.

Because after the shortcuts of looking up external materials were cut off, the remaining victory or defeat would return to the internal logic, institutional boundaries, and auditing capabilities within the data.

Jiang Lin opened the Windows command prompt window and created five categorized folders under the project directory.

He unzipped the dataset's archive package.

After the progress bar finished, four behemoths appeared inside the data_raw folder.

train_features.parquet, train_labels.csv, test_features.parquet, sample_submission.csv.

The second-hand computer he was currently using had a pathetic memory of only 8GB.

If he directly used Pandas to read this 3-some GB of Parquet-format data in full into memory, this second-hand machine would most likely freeze directly.

Thus, he opened his editor and wrote a short Python script of several dozen lines.

It only read the header metadata and the first one million rows of sample slices in the front to scout the path.

Ten minutes later, he pressed the run button.

An extremely long string of dazzling red error messages instantly flashed across the terminal.

Jiang Lin stared at that string of red characters, stunned for two seconds, and then chuckled wryly.

The Real World was just this full of dark humor.

You were full of ambition, thinking you were about to face the unfathomable capital market and tear apart complex financial microstructures with your bare hands, yet your very first slash landed solidly on a missing underlying dependency library.

"Fine then, logistics first."

Jiang Lin shook his head and typed the installation command into the command line.

Watching the bouncing download progress bars on the screen, he casually recorded the version numbers of these packages, without missing a single one, into the aforementioned audit_log.

At nine-forty-six in the evening, the environment configuration was completed.

The first batch of one-million-row sliced data was finally successfully read into memory.

More than a dozen routine market quote fields were neatly printed on the screen, followed at the end by three auxiliary fields generated by the platform itself.

session_id, source_flag, row_weight.

Jiang Lin was in no rush to look at those price numbers; instead, he stared at these last three fields for a few glances and evaluated them as:

[Whenever fields are called internal quality groupings by the platform side, do not use them as features first; audit them as potential pollution stratification variables first.]

Next, he began conducting the most basic reconnaissance audit on this one million rows of slices.

Jiang Lin translated several basic pieces of common sense in his brain into a few clean and neat aggregation codes and clicked run.

The laptop emitted a low roar, and the CPU fan began to spin at high speed, as if protesting this exploitation.

The first round of the auditing script ran for a full twenty-plus minutes.

Inside the terminal, rows of statistical results slowly popped out.

Just as Jiang Lin had anticipated, most of the records looked very smooth and normal, but there was also a small handful of records so dirty that they were unwatchable.

After all, the task of this competition was to find dirty data, so anomalies were within expectations.

Jiang Lin calmly wrote a few lines of plotting scripts, randomly drew a few of those filtered-out anomalous samples, and plotted them into trend charts.

These anomalies were all too shallow.

Writing a few if-else conditional filters would be able to catch them completely clearly.

Jiang Lin calculated in his heart that if he only relied on these surface-level business logics to submit, he should be able to steadily get an above-average score; he wouldn't get the bonus, but at least he could pass.

If the task only stopped at these surface-level logical errors, it wouldn't be worth the platform setting up manual report scoring.

Jiang Lin rubbed his somewhat sore eyes, took a sip of water, and continued reading down to the statistical results of the fourth question.

At one-twenty in the morning, when the distribution matrix of missing values rendered on the screen, he finally discovered the first anomaly that made him furrow his brows.

The missing values in the data did not appear uniformly.

If it were a general low-liquidity target, occasional data missing would be very normal.

Or alternatively, if a certain stock's exchange interface suddenly lagged, causing a gap of a few minutes, that would also make sense.

But the problem displayed on the screen now was that the missing values were extremely densely concentrated on the exact same minute_id segment across several date_id s.

Moreover, what was even more eerie was that the missingness of these minute_id s spanned across hundreds and thousands of different symbol_id s in the samples.

To see clearly, Jiang Lin plotted the missing distribution state of this one million rows of data into a two-dimensional heat map.

The horizontal axis was time, and the vertical axis was the target code; places with data were dark blue, and places with missingness were dazzling bright yellow.

Strikingly appearing on the screen were a few bright yellow strips traversing the entire map, just like a whole piece of dark blue fabric that had been cruelly slashed with a few parallel giant wounds by someone with a knife.

This wasn't a problem with a certain stock's trading; it wasn't the market behavior of the target itself!

This looked more like, during that specific time period, a certain core server responsible for collecting market data suddenly crashed, or a certain dedicated line of the underlying provider broke, causing the data of the entire market on this slice to experience full-link collective packet loss.

Jiang Lin immediately switched back to audit_log and heavily typed a line of words into the log.

[Anomaly Type A: Horizontal time cross-section-level missingness.]

[Inference: This anomaly does not belong to the trading characteristics of the market itself; it is with a high probability a failure of the data provider's underlying interface, or a timeline alignment error that occurred when the platform spliced multi-source data.]

Following this train of thought, he checked downwards, attempting to circle out all time periods where horizontal faults had occurred.

Next, he made a very small but fatal common-sense mistake.

When boxing selections on the heat map, he found that not only did certain random days have horizontal wounds, but almost in the middle section of every date_id, there was a large chunk of regular, market-wide blank areas.

Jiang Lin unhesitatingly marked this regular blank area as an anomalous missingness as well.

But at the very instant he pressed the save key, his fingers hovered mid-air.

"Wait, all targets, every single day, cut off punctually at the exact same position in the middle, and recover punctually as well?"

Jiang Lin suddenly realized something, quickly switched out of the code interface, opened the browser, and typed A-share trading hours into the search engine.

The page immediately popped out results: morning 9:30 - 11:30, afternoon 13:00 - 15:00.

There was a one-and-a-half-hour market closure at noon.

Jiang Lin slapped his own forehead.

He had mistaken this part of the regular blank space for data loss.

He opened the data dictionary and re-examined that extremely vague session_id.

Although the platform had desensitized and shifted the specific times, session_id still honestly retained the structural information of the two trading sessions each day in the morning and afternoon.

Jiang Lin immediately withdrew all those marks, returned to the log notebook, and candidly recorded his own mistake:

[Misjudgment 001: Misidentifying the Physics time breakpoint caused by the normal noon market closure as system-level data missingness.]

[Cause Analysis: Overly superstitious about data appearances, failing to establish the Physics constraint premises of the Real World market before modeling.]

[Correction Plan: All subsequent statistical and cleaning actions must first strictly segment and isolate daily market quotes by session_id for modeling, and calculating any differential features across market closure periods is prohibited.]

After writing this section, he felt a wave of relief.

This was the thrill of error correction.

Admitting ignorance, eliminating interference, and tightening the boundaries of the system once again.

Time had moved forward to two-thirty in the morning.

Jiang Lin rubbed his cheeks, dispelling a trace of sleepiness, and opened the baseline.ipynb file provided by the official.

This was a demonstration-level benchmark code written in Jupyter Notebook.

The platform side wrote it very crudely, mainly to demonstrate how to read in data, construct features, and run through the submission process.

The subsequent model part was temporarily not executed by him; he only looked down along the preprocessing code.

Jiang Lin scanned downwards line by line, his gaze rapidly parsing the logic behind these codes.

When seeing line sixty-seven, his hand holding the mouse suddenly stopped.

That was a piece of code for feature standardization.

In order to allow data of different dimensions to be trained under a single scale, it directly merged the training set and the test set into a single large table.

Then, it calculated the global mean and global standard deviation of each feature in this large table.

Finally, it used this global statistic containing all future data to scale the entire dataset.

In many online open-source tutorials or novice data analysis assignments, this writing style was very common.

Getting everything done with a single fit_transform statement—convenient, effortless, and the code also looked very neat.

But if it were in an ordinary machine learning competition that didn't emphasize chronological order, such as recognizing pictures of cats and dogs, this was called data leakage.

But in time series data like quantitative finance, it was far worse than ordinary leakage.

Because the chronological cause-and-effect order of time itself was the law of Physics of this system.

You could not use tomorrow's highest and lowest temperatures to redefine today's thermometer scale's zero point.

If you did this, today's data would inevitably be mixed with foresight into the future.

Jiang Lin did not immediately draw a conclusion in the log.

The scientific spirit demanded evidence, not a whim-based intuition.

He immediately created a concise and powerful verification script.

He only took 10 symbol_id s, and only took two date_id s closely adjacent on the timeline as the training window and test window.

He made two sets of controls—one set copying the Baseline's merging logic, and one set strictly adhering to temporal causality using only historical windows.

The script finished running in one second, and Jiang Lin brought up the line chart.

The results were extremely clear.

During the test window period when drastic price fluctuations occurred, the first group, because it had seen these drastic fluctuations in advance, caused the standardized anomalous values to be significantly suppressed.

In other words, this Baseline provided by the official was imperceptibly kneading future information into the initial preprocessing steps.

It was like a siren drugged with a knockout potion, making a portion of anomalous data that ought to have been extremely dazzling look not so anomalous after all.

This was not some participant maliciously hunting for code vulnerabilities to cheat; it looked much more like a stupid engineering oversight committed by the platform's technicians while rushing to meet a deadline.

This oversight was enough to render a portion of the scores on the public leaderboard utterly meaningless.

At the very least, those solutions that continued to follow the official baseline preprocessing pipeline were no longer clean.

Jiang Lin looked at the two diverging broken lines on his screen, and suddenly, the words Professor Zhao Mingyuan had tapped on the blackboard earlier that day in Classroom A301 at Jiangda University echoed in his mind.

"The most dangerous time for an approximation method is not when the error it calculates is large, but when the premises it relies on have already collapsed while you are still blindly using it as usual without knowing it."

That was precisely the current predicament of the baseline.

It was not that the accuracy produced by the model was insufficiently high, but that right at the very first step of preprocessing, it had already violated the causal laws of the time series from its Physics foundation.

At 3:11 AM in the dead of night, the entire city was fast asleep.

At the very end of his logbook, Jiang Lin typed out a line of text.

[Anomaly Type E: System-level future information leakage exists in the baseline.]

[Mechanism: Merging the training and testing sets during the preprocessing stage for global standardization calculations fundamentally shatters the causal structure of the time series data, causing the anomalous features of the test window to be smoothed out and diluted.]

[Note: This phenomenon is by no means a loophole that contestants should exploit to boost their scores, because a model trained on it will lead to a dead end in reality. It is perhaps a security-level flaw inherent to the competition problem itself that ought to be submitted directly to the platform for review.]

Over the next two days, Jiang Lin hardly left his room for long periods.

Aside from his mother punctually knocking on the door to call him for his three meals a day and occasionally bringing in some freshly cut fruit to place on his desk, he remained welded in front of his computer screen like a sculpture for the rest of the time.

During this time, his head teacher, Lao Liu, grew genuinely worried and gave him a call.

Upon learning that he was holed up safely at home, Lao Liu finally felt at ease.

Over these two days, Jiang Lin sliced up and precisely allocated all his time among three of the most fundamental and tedious tasks.

Establishing a rigorous Physics constraint system for trading sessions, constructing a logical consistency network among underlying fields, and building a comprehensive classification dictionary for missing data mechanisms.

He encapsulated all these audit rules into modules that could be independently toggled on and off.

Much like a precision-operated machine tool, every data stream that passed through—which cut it experienced, which alarm it triggered—was meticulously recorded in the logs.

Every ultimately outputted anomaly label could be traced back step-by-step to the fundamental Physics reason it was originally triggered.

On the morning of the third day, Jiang Lin generated his first submission.csv and clicked the submission button on the platform.

The server behind the platform spun for a few moments, and the automated scoring results quickly refreshed on the page.

Score: 0.7812, Rank: 13th.

Near the top of the page, several accounts with flashy usernames had already surged past 0.85, leaving him far behind.

Jiang Lin looked at his ranking steadily parked in the middle, then clicked to open the public submission descriptions of the top-ranking competitors.

To win prizes, contestants were required to briefly outline their methodology in the descriptions.

The text was written exceptionally well, brimming with a sense of cutting-edge academic sophistication.

Such as ensemble anomaly detection frameworks based on Isolation Forests and LightGBMs, or utilizing LSTM-AutoEncoders for unsupervised time-series anomaly recognition...

In the description of a contestant currently ranked second, Jiang Lin even saw the following sentence: "This scheme constructs a complex network with deep expressive capability, thoroughly learning the market's microstructure features and price game logic to accurately pinpoint anomalous nodes."

Jiang Lin shook his head and closed these flashy descriptions.

Using a mere 3.4 GB garbage slice sample that was heavily desensitized by humans, riddled with supplier packet loss and platform stitching errors, and where even the baseline leaked future information, to teach a deep neural network the micro-game logic of capital markets?

That was just as absurd as expecting to deduce the origin of the Big Bang by measuring a pile of garbled background noise amidst ruins filled with electromagnetic interference and rusted probes.

He proceeded to download the small batch of error sample feedback returned by the platform following his recent submission.

After analyzing the feedback, he found that the net woven purely from rigid logic had indeed missed a few covert anomalies.

Especially isolated fluctuations where a sharp price spike occurred over an extremely short time span, only to rapidly fall back to normal levels the very next minute.

These fluctuations did not violate field consistency in terms of trading volume and basic price relationships, making them very difficult to pin down using basic rules alone.

Thus, he opened his code and calmly appended a robust statistical module.

He fed the data for each symbol_id into rolling windows separated by session to calculate the MAD for each point.

After writing the MAD module and using it as a supplementary alarm system.

Jiang Lin clicked submit for the second time.

Score: 0.8105, Rank: Rose to 9th.

Still, he failed to squeeze into the frontline battle for the top spots.

But none of that mattered anymore.

Because inside his desktop folder, the analysis report accompanying the code had already reached its twentieth page.

He updated the version logs with all the model differences, shifts in methodology, and mathematical explanations regarding the statistical robustness of the MAD from his first two submissions.

At 10 PM, in the dead of night.

Jiang Lin returned to his desk holding a cup of freshly brewed strong tea and opened a blank PDF template.

He gave the report a plain title: "Minute-Level Market Data Anomaly Detection and Backtesting Pre-Audit Report".

The first page contained the report abstract.

He typed out the words that had been swirling in his mind for the past two days.

"The core viewpoint of this report is that the anomalies in this desensitized dataset do not constitute a single-dimensional statistical distribution classification problem. Through logical disassembly, the contamination sources of this data stem from at least four distinct pipelines: differential contamination caused by improperly isolated trading session boundaries, inconsistent reinstatement calibers caused by corporate actions or scaling, sporadic missing data caused by the targets' own low liquidity, and cross-sectional missingness caused by the data supply chain or platform desensitization stitching process."

"Furthermore, this report notes a significant synchronous relationship among certain horizontal missing bands, sudden price scale shifts at adjacent date_id boundaries, and changes in the platform's internal quality grouping field, source_flag. Because the true business meaning of source_flag has not been publicly disclosed, this paper does not offer a definitive interpretation of it, and merely suggests that the organizers conduct a stratified review of the underlying data sources, cleaning batches, or supplier interface switching records based on this field prior to the final evaluation."

"If contestants ignore the aforementioned data generation pipelines and directly treat all anomalies as the same type of statistical classification task to be fitted by black-box models, the models are very likely to mislearn the engineering defects of the data supply chain itself as market trading behaviors. Even if such results achieve high scores on the public leaderboard, they lack generalization value in Real World engineering scenarios."

After typing the final punctuation mark, Jiang Lin leaned back and rubbed his sore neck.

After finishing the abstract, he spent an hour on the second page using a drawing tool to create an extremely professional process topology diagram.

The title was: Data Generation and Contamination Pipeline Tracking.

Along this lengthy pipeline, beneath every node, he used red arrows to precisely mark the types of contamination that could occur—and had already occurred—on this dataset.

He wanted to use this diagram to drag the other party directly into his perspective and make them see clearly: I am not solving an algorithmic problem; I am performing surgery on your entire measurement system.

Beginning on the third page was a detailed description of the core anomaly types defined by Jiang Lin.

Jiang Lin picked up his teacup and took a sip; the walls of the cup had already gone mostly cold.

The cursor blinked on the blank page at the very end of the document.

He took a deep breath and typed out the standalone chapter title.

Class E: Baseline-Level Future Information Leakage.

Once this chapter was written in, the nature of the entire report would change.

The first four types of anomalies were all about earnestly completing the task assigned by the organizers to hunt for dirty data; no matter how deeply one found them, it was simply the duty of an exceptional contestant.

But this fifth type was equivalent to directly flipping the organizers' table.

It was tantamount to telling the judges that the official demo code they provided was rotten right down to its foundations.

If the tone was written too softly, the reviewing experts might skim right past it during a quick read and act as if they never saw it.

If it was written with too much aggressiveness, out of psychological defense to save face for the platform, the reviewers would most likely view him as a troublemaker who was forcibly nitpicking just because he couldn't get a high score, and sweep his report straight into the trash heap.

Jiang Lin leaned back against his chair and pondered for a moment, ultimately choosing an irrefutable writing style.

Mixing in no emotional commentary whatsoever, he provided nothing but ironclad runtime evidence.

He pasted in the entire minimal reproduction experiment he had prepared long in advance.

Following the detailed data and comparative charts, he wrote down the final conclusion.

"Experimental evidence demonstrates that if the test set is blindly used during the feature preprocessing stage to participate in calculating global statistics, the distribution state of the anomaly scores will be directly contaminated by future sample fluctuations. This not only logically shatters the fundamental baseline of time-series analysis, but furthermore causes the current public leaderboard scores to reflect, to a large extent, merely the ability of certain contestants to overfit using this data-splitting flaw, rather than the true anomaly detection capability of their algorithms."

At 8:17 AM on the fourth day, less than six hours remained before the submission channel for the entire challenge closed.

Jiang Lin clicked submit for the final time, uploading the third version of his model results.

The page refreshed, and the automated score came out.

Rank: 7th.

Still, he failed to break into the peak of the pyramid formed by the top five positions eligible for superficial grand prizes.

But he made no further attempt to tweak any parameters to climb the leaderboard.

Instead, he bundled the final version's clean result CSV, all the modular and heavily commented Python scripts, the requirements text documenting environment dependencies, and the twenty-four-page PDF audit report together into a zipped archive package.

He uploaded it to the final deliverables column of the system.

A blue loading circle appeared in the center of the webpage, spun for two or three seconds, and popped open a green prompt box reading Submission Successful, Thank You for Participating.

Jiang Lin let out a long breath and leaned back heavily against the chair.

As for when the other party would see it?

Once they understood it, would they frankly admit the system error, or choose to stubbornly deny it to cover up the mistake, or even outright delete his posts and ban his account?

All of these variables were no longer within his current control.

Yet on this barren slice of data, he had already pushed logic to its absolute limit.

Jiang Lin closed the browser, stretched his somewhat stiff shoulders and neck, picked up the cup on his desk, pushed open his bedroom door, and headed to the living room to pour some water.

Ten minutes later, he returned to his room holding his water cup. Just as he woke up his computer screen, the notification bell in the upper right corner of the competition webpage, which he hadn't closed, indicated that there was a new internal message.

The sender bore a blue checkmark verification: Platform Technical Group Administrator - DataOps_03.

Jiang Lin sat down and clicked open the private message.

"Hello, Contestant No. 1453. The review panel has just made a preliminary review of your attached materials. We noticed that in the Class E chapter of your report, you gravely pointed out that the baseline code might have information leakage issues during the preprocessing stage."

"Since this issue, if proven true, will seriously impact the fairness of the final evaluation of this challenge, could you please separately submit an independent minimal reproduction experiment code package to this account within 24 hours, along with a long list of rigorous submission requirements for the reproduction package."

"Postscript: Before the issue is thoroughly verified and an official announcement is issued, please temporarily refrain from publishing any derivations or screenshots concerning this vulnerability in the public discussion area of the competition. Thank you for your cooperation."

Well then.

Although the tone of this message was official and stern, the meaning between the lines was already quite straightforward—they had understood it, and the entire technical team must not be feeling relaxed at this moment.

Because once this issue was completely verified, it would constitute a severe, accident-level problem-setting error.

If Jiang Lin were a hothead who directly threw the report into the open-source communities and quantitative forums across the entire web, the entire technical team behind the platform would be reduced to the laughingstock of the industry.

Jiang Lin unhurriedly placed his water cup on the corner of his desk, reopened the command-line terminal, and created a new folder alongside the original project directory.

mkdir minimal_reproduce_leakage

Then, he skillfully switched into the audit_log text file that had accompanied him for the past four days.

Moving his cursor to the very end of the file, he typed out a line recording a change in system status.

[Phased event triggered: Although the initial final submission failed to claim the artificially inflated top spot, the core report successfully breached the opponent's defenses, triggering the manual review process by the competition's technical group and data provider.]

[Next execution objective: Construct a clean, minimal reproduction experiment within two hours, using code to force the opponent's engineering vulnerability to move beyond a one-sided opinion and solidify it into an irrefutable, runnable physical fact.]