Stringers
Luciferbobo/KiT · Luciferbobo

"Kronos got beaten by an Alibaba researcher": how a K-line model earned stars with one comparison table

What it is An open-source model released as a side project by an Alibaba video-AI researcher. It uses video-generation methods to draw an entire future stretch of K-lines for stocks and crypto in one pass. Weights and demo data are free, and it runs on a 6GB GPU.

Why it took off It picked Tsinghua's 48k-star Kronos as its target, and its comparison table claims it wins on every timeframe. On Nov 9 and 12, two Chinese and English X recommendation accounts with about 10k followers each reposted it, following the README. Daily star gains jumped from around 4 to 20–40.

CHow you could follow suit Don't build another K-line model. Apply its launch playbook to your own paper or project: go after a leader with tens of thousands of stars, put the comparison table on the first screen, release weights, data and a one-click script on day one, and aim for recommendation accounts around 10k followers for the first wave. If you want a piece of the finance hype, build a neutral third-party benchmark leaderboard instead.

C grade: the playbook is clear and worth learning, but the niche is dominated by Kronos, and latecomers mostly stall at a hundred or two hundred stars. KiT itself has only 266. Building a model that can challenge the leader takes research ability and a lot of compute, and the stars it earns don't justify that. Chatter in the small circle
Data (updated daily) · Repo created 2026-09-28 · 348 stars · 7 days +147
GitHub · stars
020040010-1110-2110-3111-0911-20348
Data 2026-10-11 → 2026-11-20 (41 days, 40 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

KiT is an open-source repo attached to an arXiv paper, doing K-line prediction for stocks and crypto. It picked up 266 stars in 49 days, not a big number, but the method is clear: find a leader with 48k stars, put "it wins on every timeframe" on the first screen of the README, release the weights, data and scripts in full, then wait for recommendation accounts to repost it. Below: what it is, where the stars came from, who can copy this playbook, and why I wouldn't recommend building yet another K-line model.

A post with 9,000 views

On 2026 9 15:57 (UTC), an English X account with 15.7K followers, @phosphenq, posted: "Kronos (Tsinghua's 47K-star trading AI) - F*CKED by Alibaba researcher..." The post goes on to say he works on video AI by day and quietly released a stock model that beats Kronos on every timeframe from 1-minute bars to daily, and it's free. Next comes a 7-step tutorial. Step one: "open Claude Code in an empty folder (needs an Nvidia GPU with 6GB+)." Step two is a line you can paste straight in, telling Claude Code to clone the repo from GitHub and pull the model and data from Hugging Face. The last step: "want proof? run: python eval/cal_metrics.py"

The post has 9,201 views, 173 likes and 253 saves. More saves than likes suggests many people plan to come back and actually run it (X page data, 11-16). In the month before, KiT gained only about 4 stars a day. On the data day after the post (11-11), it gained 30 in one day.

Three days later, on Nov 12, a Chinese GitHub-recommendation account @_zheergen with about 9,100 followers posted a 67-second video: "No way! Someone turned stock K-line prediction into an AI foundation model." It listed six highlights, nearly all copied from the README. That post got 8,628 views, and KiT gained another 17 stars the next day.

What it is

KiT stands for K-line Diffusion Transformer. The author, Boyu Zhang, lists on his personal page that since March 2026 he has been a senior research engineer at Alibaba working on video-generation foundation models, previously at SenseTime. His published papers are in NeRF and rendering, not finance. In October 2025 he forked Kronos, and he has also translated the Chinese edition of the "Quantitative Finance Interview Green Book".

The model applies video-generation methods to K-lines. You give it a stretch of historical K-lines, and it generates an entire future stretch in one pass, producing 16 paths to show the possible range. Kronos writes one bar at a time, so an early error compounds into bigger errors later. KiT uses that as its selling point. Training data covers US stocks, A-shares and crypto across seven timeframes, from 1-minute bars to daily, totaling billions of K-lines.

What was released: 284M-parameter weights, demo data for 100 stocks, single-stock prediction, validation inference, and a backtest script with fees and slippage. It runs on a 6GB GPU. Training code has not been released. Weights went up on Hugging Face in October 2. By 11-16 the model had 1,963 downloads and the demo dataset 835 (Hugging Face page). It has a few hundred stars but nearly two thousand downloads, which suggests people are actually running it, not just bookmarking it.

How the stars came in

TimelineWhat happenedstar
09-28Repo created, arXiv paper posted the same day0
09-29A Hugging Face staffer opened an issue, the routine invitation to upload the paper page and weights
10-02Inference code, weights and demo data released together
10-11Paper aggregator sites and quant-paper weeklies started listing it20
10-16A WeChat group QR code added to the README
11-10Around 4 a day through the month, with no spike148
11-09 → 11-11@phosphenq's English post178(+30)
11-12 → 11-13@_zheergen's Chinese video post201
11-13A collaborator submitted a PR adding an A-share live simulation page, with a "Live Prediction" badge at the top of the README
11-16+39 in one day, source not found266

(Star counts are daily GitHub snapshots. The last row is the GitHub page as of 11-16.)

The first month's 4 a day was the natural traffic of an ordinary arXiv paper. The inflection point lines up exactly with the two X posts. It never made GitHub trending (hard for a repo less than two months old), and I found no coverage on Hacker News, Reddit, Zhihu, Bilibili or WeChat Official Accounts. The +39 on 11-16 was the biggest day, but I can't find a matching post. It may have come from a WeChat group or a Chinese channel I couldn't find (inference).

The simulated live-trading page went up in Nov 13. It says it rebalances weekly on KiT's predictions, buying the top 10, with cumulative NAV at +35.6%, versus CSI 500 at −13.68% over the same period. The page itself notes that the model's training cutoff date was never verified, so the result can't be treated as out-of-sample. The tutorial says the model only learned data through the end of 2025, but the second weight file is named with "train260410", which suggests the training data runs to April 2026 (inference). Either way, the page went live after the two posts, so it's an amplifier, not the cause.

Why it took off

1. It rides the Kronos hype. Kronos is a K-line foundation model from a Tsinghua team, 11-16 has 47,875 stars, and Chinese big accounts and Zhihu and Bilibili have written about it a lot. Kronos got hot again in early November, gaining 691 stars in 11-03 alone. The second and third rows of KiT's main README table are Kronos and the fine-tuned Kronos, so people following Kronos who want "is there something stronger?" find it with one search.

2. It handed recommendation accounts ready-made headlines. "Beats Kronos in every timeframe", "volatility forecasting 0.66 vs 0.50", plus the contrast of "a big-tech video AI researcher built this on the side, free". Recommendation accounts could write a post almost by copying the README's first screen. Both posts drew their content almost entirely from it.

A caveat: in the returns table, all seven of KiT's columns are bolded, and it does beat Kronos on every column. But in the same table, Chronos-Bolt is higher on the 1-minute, 30-minute and 2-hour columns, and TimesFM 2.5 is higher on the 30-minute and 2-hour columns (GitHub README, 11-16). KiT's average return is only 0.005 higher than fine-tuned Kronos. It's a single-author preprint, not peer-reviewed, and no third party has reproduced it.

3. It runs out of the box. The recommendation posts could confidently say "try it as soon as possible" because the weights, data and scripts are all there, and you can hand the setup to Claude Code in one sentence. FinCast, a financial time-series model that also has a paper (it placed at CIKM 2025), has only 143 stars. ZJ_Quant_03, which does A-share K-line modeling, has 99. Neither has KiT's level of completeness, and no one is promoting them.

The demand is also real. In Oct 15, someone opened an all-caps issue in the repo: "I BEG U ON MY KNEES: PLS RELEASE THE TRAINING CODE". They said they had only 105 euros in the bank and an old 1070 Ti, couldn't afford an H200, and that same day they published their own training and fine-tuning code rebuilt from the paper and weights. The author replied the next day that the training code was still being cleaned up and the "target is to release it in about a month". As of Nov 16, it still hasn't been released.

Who it's competing with

ProjectWhat it isstars (around 11-16)
Kronos (Tsinghua)The leader among K-line foundation models, released 2025-0747,875
TimesFM(Google)General-purpose time-series model35,563
Chronos / Chronos-2(Amazon)General-purpose time-series model6,309
Timestral(Mistral)General time-series model released by 10-151,197, more than 4× KiT's gain in half a month
Kronos side repositoryWraps Kronos in a UI, writes Chinese tutorials427 / 152 / 73
FinCastFinancial time-series model, has a paper143
ZJ_Quant_03A-share K-line model99

In this niche, big-tech general models get stars on name alone, Kronos dominates, and latecomers mostly stall at a hundred or two. KiT is one of the few latecomers that took off, and it still has only 0.5% of Kronos's stars. No one has built a second one following KiT yet.

If it were you

C: The launch playbook is worth learning, but building another K-line model isn't worth it. A latecomer's ceiling is around a few hundred stars. KiT took 49 days and two recommendation posts to reach 266. Building a model that can beat Kronos takes research capacity and a lot of compute (the H200 the issue author mentioned). Take away the playbook, not the niche.

Good fit for: researchers and grad students who already have a paper to publish, in any field. Applying this playbook to your own paper is far cheaper than building a project from scratch.

Bad fit for: indie devs who want to chase stars with a "stock prediction model". No compute, a crowded niche, and a comparison table that falls apart when third parties can't reproduce it. Financial prediction also draws "another pump-and-dump" skepticism.

How to do it:

  • Pick a niche led by a repo with tens of thousands of stars, build something aimed squarely at it, and put a comparison table on the first screen of the README so recommendation accounts can write "X got destroyed".
  • Release the paper, weights, demo data and a directly runnable inference script all on day one. State the hardware requirement clearly (KiT says a 6GB GPU), and add a one-line setup prompt that can be pasted into Claude Code. Completeness determines whether recommendation accounts dare to push it.
  • Add a page that's updated continuously so people come back each week (KiT has a live-trading simulation page), plus a community entry point.
  • If you want to ride the finance wave, instead of training another model, make the comparison table nobody has made: run Kronos, KiT, TimesFM and Chronos on the same fresh dataset and publish a continuously updated, third-party benchmark leaderboard. (Inference: right now every lab's comparison table is self-reported, so a neutral table is itself recommendation material.)

How to get users: Hugging Face paper pages and quant-paper newsletters only bring a few stars a day. The real inflection point is Chinese and English GitHub project recommendation accounts on X with around 10k followers. Use the "X got destroyed" style in English and the "Holy cow! Someone just built..." style in Chinese, with content taken straight from your README's first screen. So write the first screen in a way they can copy.

When to move: anyone building tooling around KiT (training, fine-tuning) needs to beat the official training code release. The author said mid-November or so. Anyone building a benchmark should launch while the Kronos and KiT wave is still going, before December.

Signs you misread it: (1) two weeks later, KiT is back to single-digit daily stars, meaning this was a one-off bump from two posts, and the value of copying the playbook drops. (2) A third party reproduces it and KiT can't beat Kronos. The "take down the leader" narrative would backfire. (3) The reverse: if KiT passes 1,000 stars before December and people start building a second one, the latecomer spot is worth more than it looks, and the financial time-series niche can still fit one more entrant.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report
Competitors
NameWhereLaunchedScaleHow it differs from this one
Kronos (Tsinghua shiyu-coder/Kronos)Open source2025-0747,858 stars; first 7 days +2,106,11-03; single-day +691; covered heavily by GitHubDaily, Zhihu and BilibiliThe leader. It discretizes into tokens and predicts autoregressively, one bar at a time. KiT generates a whole future path at once with a probability band, and the whole narrative is "beats Kronos".
TimesFM(Google)Open source2024-0435,563 starGeneral-purpose time-series foundation model, not built for K-lines. KiT's table uses it as a baseline.
Chronos / Chronos-2(Amazon)Open source2024-026,309 starGeneral-purpose time-series model
Timestral(Mistral)Open source2026-10-151,197 starGeneral multivariate time-series model from big tech. Its star count in half a month is more than 4× KiT's.
FinCast(vincent05r/FinCast-fts)Open source2025-08143 starA financial time-series foundation model, also placed at CIKM, but it didn't pitch "beats Kronos" and no one promoted it.
ZJ_Quant_03 (FSQ K-line foundation model)Open source2026-0599 starA K-line model for A-share stock selection, with no paper and no buzz.
Sundial / Timer-XL / TiRex-2 / MOMENT / Lag-LlamaOpen source2024–2026Went from 150 to 1,600 stars and basically stalledAcademic general-purpose time-series model
facecat-kronos and other Kronos derivative librariesOpen sourceFrom ¥2025-08427 stars, etc.Wrap Kronos in a UI or write tutorials. Not new models.
microsoft/MarSOpen source2024-101,888 starUses a generative model for market simulation, not prediction
Where users come from
  • Verified arXiv and paper aggregators, HF: 09-28 posted to arXiv. It's indexed on ML-Quant, ml4trading and HF Papers. NielsRogge on HF opened an issue inviting an upload, and 09-29 10-02 uploaded the weights and code to HF. In the month before, it gained about 4 stars a day.
  • Verified X English recommendation account @phosphenq: 11-09 posted "Kronos…F*CKED by Alibaba researcher" with a one-click Claude Code install tutorial. 9,201 views, 173 likes, 253 bookmarks. On the next data day, 11-11 +30 (versus about 4 a day before).
  • Verified X Chinese GitHub recommendation account @_zheergen: 11-12 posted a video, "Holy cow! Someone just built stock K-line prediction into an AI foundation model", with 8,626 views. 11-13 +17.
  • Inference Live-trading page + WeChat group (private traffic): 10-16 added a WeChat group QR code to the README. 11-13 a collaborator's PR launched trading.etfnav.org (NAV +35.6% vs CSI 500 −13.68%), and a Live Prediction badge was added at the top of the README. 11-16 +39, but I couldn't find a matching post.
  • Verified quant-circle X roundup: @LeeLeepenkman's 10-20 paper-roundup tweet listed it, small volume

Why KiT (Luciferbobo/KiT) took off

Channel: GitHub. Data card refreshed at 2026-11-16; the GitHub API was queried live at 11-16. Data sources: our github and hackernews databases, the GitHub API, the Hugging Face API. Web search uses Jina / Exa / Brave; tweet data uses fxtwitter (read via Jina).

One line

KiT is an arXiv paper (2609.34507,2026-09-28) with an open-source repo. The author is a video-generation researcher at Alibaba. The paper proposes a "K-line diffusion Transformer" foundation model that predicts OHLCV for stocks and crypto. It uses flow matching to generate multiple future OHLCV paths at once, rather than writing bars one after another like Kronos. The README has a table headed "beats Kronos / TimesFM / Chronos in every timeframe". The weights, a demo dataset of 100 stock(s), inference code and backtest code are all public and free, and it runs on a 6GB GPU.

Its stars largely ride on Kronos's heat (Tsinghua's K-line model, 48k stars). Kronos spiked again in early November, and two mid-sized X accounts posted about it in sequence: @phosphenq in English, "Kronos got destroyed by an Alibaba researcher", and @_zheergen in Chinese with a video post. The inflection in the star curve came right after those two posts. The total is still small, 49 stars in 265 days. It has just surfaced, not broken out.

1. Basic data

  • Created 2026-09-28, and the paper went to arXiv the same day. The data card recorded 11-16 with 247 stars and 27 forks. The GitHub API now shows 265 stars, 31 forks and 8 watchers.
  • Daily curve (github database; data from 10-11 to 10-27 comes from deeply's historical data, snapshots taken once a day from 10-28):
    • 09-28 → 10-11: rose to 20 stars, with almost no one looking in the first two weeks.
    • 10-11 → 11-10:20 → 148, averaging about 4 a day, with no peak.
    • 11-11 itself: +30,11-13 +17,11-16 +39. Over the last 7 days: +103 in total; the first 7 days were +34 (data card).
  • It never hit trending. Repos less than about 60 days old can't reach trending anyway (per the observed results in the github database SCHEMA), so this doesn't tell us much.
  • Hugging Face (API, queried live at 11-16): the model Lucifer744/KiT_model has 1,963 downloads and 4 likes, and the demo dataset has 835 downloads. Both were created 10-02. Stars are only 265, yet model downloads are close to 2,000, so people really are downloading the weights and running them, not just starring it to bookmark.
  • Code status. The data card says "zero code", which is wrong. The commit "add complete project code" from 10-02 has already added the inference code in kit_v1/ (DiT, RoPE, AdaLN, sampler), eval/ (single-stock prediction, validation-set inference, metric calculation, backtesting), and configs for three model sizes. The only thing not released is the training code; the README says "Train: TBD".
  • Author Boyu Zhang (GitHub Luciferbobo, 104 followers). His homepage thatbobo.com says: since 2026-03 a Senior Research Engineer at Alibaba, working on video generation foundation models; before that at SenseTime and Jianyi Robotics. MS from UCLA, BS from Southeast University. His papers are all in visual areas like NeRF and rendering, finance is not his field. He also forked Kronos (Luciferbobo/Kronos-finetune, 2025-10) and translated the Chinese edition of the Quant Finance Interview Green Book (28 stars). His other two high-star repos are Awesome-One-Step-Generation (163) and vibe_coding_sim (94).
  • The paper's second author is listed as H Li (Google Scholar). 11-13 The haoruilee account (Li Haorui, homepage lihaorui.com, bio says "ex @alibaba, now @QingCheng-AI") who opened a "Add live predict demo" PR to the repo is probably him, but unverified. The live simulation page trading.etfnav.org was added by this PR.

2. Competitors: who else is building "financial K-line / time-series prediction foundation models"?

NameFormLaunchedScaleHow it differs from KiT
Kronos (shiyu-coder/Kronos, Tsinghua)Open source + paper + HF weights2025-0747,858 stars; last 7 days +525, first 7 days +2,106,11-03, single day +691 (GitHub repo); GitHubDaily and other large accounts, Zhihu, Bilibili, Tencent Cloud Community have written a lot about itThe leader in this exact thing: discretize K-lines into tokens, then autoregress. KiT's whole narrative is "beat Kronos".
Kronos derivative repos (facecat-kronos, kronos-guide-cn, Kronos-demo, etc.)Open sourceFrom ¥2025-08427 / 73 / 152 starBuilding UIs or tutorials for Kronos, showing Kronos grew a user base
TimesFM(Google)Open-source general time-series foundation model2024-0435,563 stars, +244 in the last 7 daysGeneral time-series model, not built specifically for K-lines; KiT's table uses it as a baseline
Chronos / Chronos-2(Amazon)Open-source, general2024-026,309 starSame as above
Timestral(Mistral)Open-source, general2026-10-151,197 stars, 10-20 gained 202 in one dayA big-tech general time-series model released recently that got over 4x KiT's stars in half a month
Sundial / Timer-XL (Tsinghua thuml), TiRex-2, MOMENT, Lag-Llama, uni2tsOpen-source, general2024–2026150 to 1,600 starsAcademic general time-series models; growth has basically stalled
FinCast(vincent05r/FinCast-fts)Open source, CIKM 20252025-08143 starAlso a financial time-series foundation model, its paper was accepted too, but it didn't frame itself as "beat Kronos" and nobody pushed it
ZJ_Quant_03 (FSQ K-line foundation model)Open source2026-0599 starAlso a K-line foundation model for A-share stock selection, no paper, nobody spread it
microsoft/MarSOpen source2024-101,888 starUses generative foundation models for market simulation; similar direction
ByteDance ChatTS-14BOpen-source LLM2025Not verifiedUses large language models to "read" K-lines; different route

Bottom line: this track isn't short on competitors, and there's one absolute leader, Kronos. Financial time-series models that entered after it have almost none of the stars (FinCast 143, ZJ_Quant 99). KiT is one of the few among the second wave to take off.

3. Where users came from (by time)

  1. arXiv and paper aggregators (since 09-28): ML-Quant (09-29 listings, weekly ranking), ml4trading.io, mlsanity, catalyzex, Semantic Scholar and HF Papers all picked it up. On 09-29, HF's Niels Rogge opened issue #1 "Release KiT on Hugging Face", the routine HF invitation to paper authors to upload. The author 10-02 did it: weights and demo data went up on HF, and the code was released the same day. This stretch brought the first 20 to 80 stars, roughly 4 a day.
  2. Retweets from quant research circles: @LeeLeepenkman posted a roundup tweet of quant papers on 10-20, with KiT on the list. Small reach.
  3. WeChat groups (since 10-16): the README added a "KiT WeChat Group" badge and group QR code, and when the code expires you add the assistant's WeChat to get in. 11-13 switched to a new code. The etfnav demo page also says "Contact the authors / join the group". This step moves Chinese quant retail users into a private community; whether it counts as a growth channel is not yet clear.
  4. @phosphenq's English X post (11-09 15:57 UTC): the account has 15.7k followers. The post got 9,201 views, 173 likes, 253 bookmarks (more bookmarks than likes), and 21 reposts (fxtwitter). Its copy reads "Kronos (Tsinghua's 47K-star trading AI) - F*CKED by Alibaba researcher… his day job is video AI… beats it on every timeframe… for free", followed by a 7-step tutorial "install in one line with Claude Code", and the last step is "want proof? run cal_metrics.py". The next data day 11-11 is +30 (about 4 per day before that). The moment we took the snapshot and the tweet's timezone may be off by a day, so here it is matched as "within 1 to 2 days of the post".
  5. @_zheergen ("Ace, who loves pickled chicken feet") Chinese X video post (11-12 00:55 UTC): the account has 9.1k followers, 8,626 views, 64 likes, 69 bookmarks (fxtwitter). This is the format of GitHub project recommendation accounts like "Wow! Someone turned stock K-line prediction into an AI foundation model", and the same account's previous post was about a TradingView indicator library. Corresponding to 11-12 +6 and 11-13 +17.
  6. Live simulation page trading.etfnav.org (launched 11-13): an A-share simulated portfolio that rebalances weekly per KiT's predictions, buying the top 10. The page states cumulative NAV +35.60%, while the CSI 500 over the same period is -13.68%, Sharpe 1.35, max drawdown -20.51%, two-way turnover 71.9x. The page itself notes "the training cutoff date was not verified from the weights, so the results should not be treated as out-of-sample". This is a hook for retail investors, and the README header also carries a "Live Prediction" badge. Three days after launch 11-16 rose +39, but no post or article matching that day was found; the reason is unknown.
  7. Channels with no findings: HN (not in the hackernews library), Reddit, Chinese media or WeChat Official Accounts, Zhihu, and Bilibili had no coverage of KiT; the big account GitHubDaily wrote about Kronos but not KiT.

4. Why it took off

  • Borrowing Kronos's name and heat: the README's main table puts Kronos-base and the fine-tuned Kronos-base-FT in rows two and three. The phosphen post title is literally "Kronos is no longer #1". In early November Kronos itself heated up again (11-03 single day +691, first 7 days +2,106), and people following Kronos came over looking for "anything stronger".
  • A conclusion that's easy to retell: "beats Kronos in every cycle", "volatility RankIC 0.66 vs 0.50", paired with the contrast persona "Alibaba researcher, day job in video AI, built this on the side, free". Recommendation accounts nearly copy the README and can produce a viral-format post.
  • Runs out of the box: 284M weights, demo data, inference and backtest scripts all included, 6GB of VRAM is enough, and it can be installed by Claude Code in one line. So recommendation accounts dare to write "try it as soon as possible", and HF really got about 2,000 downloads. Similar projects like FinCast and ZJ_Quant lack this fully runnable setup, and nobody pushed them either.
  • Hooks for retail investors: the +35.6% NAV on the live simulation page and the weekly rebalancing signals, plus the WeChat group, appeal to Chinese quant retail users. But it only launched 11-13 and came after the two X posts, so it's an amplifier, not the cause.
  • Keep the scale in perspective: 49 days, 265 stars; next to Kronos it's a rounding error and far below Mistral's Timestral (1,197). It's just in the early phase of accelerating growth, where two posts with about 9k views each were enough to push it, which shows the base is very low.

5. Differentiation

  • Technically: KiT uses diffusion / flow matching to generate whole paths rather than autoregressively predicting bar by bar, so there is no error accumulation, and it can output probability intervals (a distribution over 16 paths). It predicts both returns and volatility, pretrained across three markets and seven timeframes (1 minutes to daily). The paper reports a mean return RankIC of 0.057 versus 0.052 for Kronos-FT; volatility RankIC of 0.66 versus 0.50 for Kronos-FT. Note this is a single-author preprint, not peer-reviewed, with no third-party replication.
  • Is differentiation why it took off? Partly. The recommendation accounts retold exactly this layer of difference, "generates whole paths, gives intervals, beats Kronos in every cycle", so the differentiation was treated as material to spread. But what made it visible was the demand Kronos had already built and the two X posts. Similar projects nobody pushed (FinCast also has a paper and also beats general models) didn't take off. KiT's return RankIC is only 0.005 above Kronos-FT, a very thin edge.

6. Copycats and visibility

  • Copycats: no project copying KiT was found. Someone in the community posted a complete training and fine-tuning codebase "rebuilt from the paper and weights" in issue #5, and in issue #4 someone begged the author to release the training code. These are demand signals, not copying.
  • Visibility: spread only within small circles. Just two X posts with about 9k views each, one quant weekly roundup, and one WeChat group; no media, HN, or big accounts wrote about it, and it never made trending.

7. What others can learn

  • Find an established leader with tens of thousands of stars (Kronos), build a model or library that is explicitly "against it", and put the comparison table on the first screen of the README so recommendation accounts can use "X got beaten" as a headline.
  • Ship the paper, HF weights, and runnable inference scripts together, plus a "one line to have Claude Code install it" prompt. Completeness determines whether recommendation accounts dare to push it.
  • Add a continuously updated "live / simulated portfolio" page so retail users come back every week; use a WeChat group to build a private community.
  • Do the first wave with Chinese and English GitHub project recommendation accounts; an account with around 10k followers is enough to push a cold repo.

8. Still unclear

  • Who brought the +39 on day 11-16 is unknown (possibly the WeChat group, an official account, or Xiaohongshu; not verified).
  • Unknown whether phosphen and zheergen posted on their own or were contacted by the author or a collaborator.
  • Unverified whether haoruilee is the second author and who owns etfnav.org.
  • The start date of the live page's +35.6% and its out-of-sample status are unverified; the page itself says so.
  • When training code will be released and whether the paper gets accepted are uncertain; if the Kronos team ships a new version, KiT's narrative loses its footing.
  • How the snapshot timing matches the tweet timing (off by 0 or 1 days) can only be roughly aligned.

Sources

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.