Skip to content

Paper map: three difficulties and three answers ​

This chapter walks you through the KiT paper from the beginning (arXiv:2609.34507). In the previous chapters you already got the model running and trained it once; now look back: why is each design the way it is, and what evidence does the paper offer that it works.

After finishing this section, you will be able to sketch the skeleton of the paper on a single page and know what questions to bring to each subsequent section.

The order for reading a machine learning paper ​

Few people read from the first line to the last. An efficient order:

  1. The abstract and the last paragraph of the introduction: what problem it solves and what it claims to have achieved
  2. The method figure (for KiT, Figure 1): how data flows through the model
  3. The main results tables (Table 1, Table 2): better than whom, and by how much
  4. Ablation experiments (Appendix E): how much performance drops when a design is removed — this is the evidence for "whether each design is useful"
  5. Then return to the method's formulas and the details in the appendix

The following sections follow this approach, with one section corresponding to one part of the paper.

The problem: why candlestick prediction is hard ​

The paper's introduction lists three difficulties:

Look it upChallenges

Click one to see.

Three answers ​

The paper's core claim is that these three difficulties should be solved together, by changing three things respectively.

What to changeKiT's approachMain responseWhich section of this chapter
What to generateInstead of predicting a single value, use flow matching to generate an entire future path at once, from which many samples can be drawndifficulty 3 (distribution), and it also avoids autoregressive error accumulationSection 2, Section 5
How to representConvert each candlestick into five logarithmic ratios: reversible, independent of price scale, and guaranteed to keep the candlestick validdifficulty 2 (non-uniformity)Section 3
Where conditions enterConditions that stay constant across the window go through AdaLN, while conditions that change per candlestick are added to the tokens; history is not noised and is placed in the same sequence as the futuredifficulties 1 and 2: letting the model fully use information like "which instrument is this, what time is it now"Section 4

Then experiments are used to show that across three markets and seven granularities, both return and volatility predictions beat existing methods; backtesting then shows these predictions are economically meaningful.

The paper's three contributions ​

The three from the original text, in plain language:

  1. The first diffusion-based candlestick foundation model, reframing prediction as "conditional path generation," which avoids autoregressive error accumulation at the root.
  2. A representation designed specifically for candlesticks (recoverable to OHLCV), robust normalization, and multi-source condition injection, allowing a single model to span markets, instruments, and granularities; no extra tokenizer is needed, so there is no tokenization error.
  3. Pretrained on billions of candlesticks across three parameter scales, achieving state-of-the-art results on all metrics.

Keep this table in mind as you read on: for each section you read, ask yourself which difficulty this design responds to, and which set of numbers the paper uses to prove it works.

Quiz timeWhich difficulty in the paper does ”stock return IC rarely exceeds 0.1” correspond to?

Check yourself ​

  • I can state the three difficulties
  • I can map "what to generate, how to represent, where conditions enter" to the difficulties
  • I know that next I should look at Figure 1, Table 1/2, and then the ablations in Appendix E

Next section: Preliminaries: from autoregression to flow matching.

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.