Skip to content

What Is KiT ​

After reading this section, you will be able to explain in your own words: what KiT takes in, what it puts out, and why it gives you "a handful of possibilities" rather than a single price.

First, Look at What It Puts Out ​

Ask someone "what will happen to this stock next week," and a sensible answer is "most likely a small rise, but it could also fall 5%." KiT's answer has the same shape: it draws 16 possible paths, from which you can read the median, the range, and how large the volatility is.

Watch onceHow KiT draws the next 20 candlesticksReal output: 300733.SZ · 1-hour bars
Anchor 2026-02-09 15:0029.37

The model first reads history: the past 400 1-hour bars (only the last 48 are shown here).

Above is a real prediction made by the official model for 300733.SZ: it read the 400 1-hour candlesticks before 2026-2-9 and drew the following 20. The median of the endpoints of the 16 paths was +3.37%, while the actual result was +4.95%, falling within the predicted range.

What It Reads In: One Candlestick Becomes Five Numbers ​

The model does not directly look at "open at 10.2 yuan, close at 10.6 yuan." Prices vary in magnitude: put a 5-yuan stock next to a 2000-yuan stock, and the numbers differ by a factor of hundreds.

KiT replaces each candlestick with five ratios (log-transformed): the gap, the body, the upper shadow, the lower shadow, and volume relative to average volume. Drag the candlestick below:

Drag it yourselfOne candlestick, turned into five numbers
Previous close
r_gapGapToday's open vs. yesterday's close+2.0%
r_bodyBodyClose vs. open+3.8%
r_upUpper shadowHigh vs. top of the body+2.8%
r_dnLower shadowBody bottom is lower than the low+2.0%
vVolumeVolume relative to its recent average+40.5%

All five numbers are logarithms of "ratios". Upper and lower shadows are always ≥ 0, so the generated candlesticks will never have a high lower than the close.

This approach has three benefits, all documented in the paper:

  1. It can be inverted: given yesterday's closing price, the five numbers let you work back step by step to the open, high, low, close, and volume—no information is lost.
  2. It ignores absolute prices: multiply all prices by 100 and the five numbers stay the same, so one model can learn thousands of stocks and several markets at once.
  3. Generated candlesticks are always valid: the upper and lower shadows are numbers ≥ 0, and a simple clipping after generation prevents weird candlesticks like "the high is below the close."

How It Draws the Future: "Flowing" Out of Noise ​

KiT belongs to the diffusion model family, specifically a method called flow matching. The process of drawing those 20 future candlesticks goes like this:

  1. First, place a blob of random noise in the future positions;
  2. The model looks at the history and decides "which way this noise should be pushed";
  3. Push a small step, look again, push again. By default it pushes 32 steps;
  4. When the pushing is done, the noise has become a plausible price path. Start over with a different blob of noise, and you get another path.

So the 16 paths come from 16 different blobs of noise, each pushed 32 steps. The historical part stays exactly as is; the model only generates the future segment.

Line them upWhat steps does KiT take, in order, to produce one prediction?
  1. 1?
  2. 2?
  3. 3?
  4. 4?

Click them one by one, in order.

What It Was Trained On ​

The training data in the paper: 3914 A-share stocks, 888 US stocks, and 28 crypto instruments, from 2016-8 to 2026-4, across seven granularities (1-minute, 5-minute, 15-minute, 30-minute, 1-hour, 2-hour, 1-day). The same design was trained at three sizes: 31000K, 1.01 billion, and 2.84 billion parameters; the largest one is what was open-sourced.

There are two open-source checkpoints:

CheckpointData usedCoverage
step_market_284M.ptTraining set up to 2023-12-31, with 2024–2025 for validation; from 2026 onward is the test periodA-shares, US stocks, crypto—this is the one used for the paper's results
step_market_284M_train260410.ptTrained up to 2026-03-09, with 2026-03-10 to 04-10 for validationA-shares only

The dates in the first row come from Section 4.1 of the paper; the model card lists its cutoff date as 2025-12-31, referring to the data used for training plus validation.

Quiz timeHow can KiT use the same model to predict both ¥5 and ¥2000 stocks?

Check Yourself ​

  • I can state that KiT's output is "multiple possible paths," not a single price
  • I know that one candlestick becomes five ratio numbers
  • I know that the "16 paths" come from 16 different blobs of noise

Next section: Set up the environment, weights, and data.

gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.