Kronos K-line model local operation: environment installation, model selection, prediction examples and backtesting pitfalls

This tutorial by Kronos focuses on the specific task in the title. Kronos is a financial K-line modeling project; the forecast output cannot be regarded as investment advice. Data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

Below is a demonstration using a test environment and revocable data. When accounts, keys, machine control, investments or emergency communications are involved, manual confirmation must be done.

First clarify whether the task is to reproduce or research

This section only deals with “first clarify whether the task is to reproduce or research”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

git clone https://github.com/shiyu-coder/Kronos.git

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Fixed Python, CUDA and dependency versions

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1
2

python --version
nvidia-smi

The completion standard is to “fix Python, CUDA and dependent versions” to obtain verifiable results, rather than the command just not reporting an error.

Create an environment isolated from other projects

1

python -m venv .venv

When a stop signal appears, resume first and do not continue to the next step.

Model weight and code version must match

This section only deals with “model weights and code versions must match”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

git rev-parse HEAD

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

The order of K-line fields cannot be guessed.

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1

# 检查 timestamp/open/high/low/close/volume

The completion standard is that “the order of the K-line fields cannot be guessed” to obtain a verifiable result, rather than that the command happens to not report an error.

Time zone, suspension and missing values are processed first

1

# 输出时间范围、时区与缺失统计

When a stop signal appears, resume first and do not continue to the next step.

What will change before and after reinstatement?

This section only deals with “what will change before and after reinstatement”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

# 保存复权方式和数据供应商

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Only run the official minimum sample for the first time

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1

python --help

The completion standard is to “run only the official minimum sample for the first time” to obtain verifiable results, rather than the command just not reporting an error.

If there is insufficient video memory, shorten the sequence and batch first.

1

nvidia-smi

When a stop signal appears, resume first and do not continue to the next step.

The training set cannot see future data

This section only deals with “the training set cannot see future data”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

# 按时间切分 train/validation/test

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Rolling backtesting is better than a random split

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1

# 使用 walk-forward 时间窗口

The completion standard is that “rolling backtesting is better than a random split” to obtain verifiable results, not that the command happens to report no errors.

Add fees and slippage to results

1

# 分别报告毛收益与净收益

When a stop signal appears, resume first and do not continue to the next step.

Compare simple benchmarks instead of just looking at models

This section only deals with “comparing simple benchmarks rather than just looking at models”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

# 与持有策略和移动平均基准比较

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Save random seeds and data snapshots

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1

git status --short

The completion standard is to “save the random seed and data snapshot” to obtain verifiable results, rather than the command just not reporting an error.

Reasons why the prediction direction is correct but unable to trade

1

# 检查换手、成交量与延迟

When a stop signal appears, resume first and do not continue to the next step.

How should research conclusions be expressed?

This section only deals with “how research conclusions should be expressed.” Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

# 只报告样本、方法、限制和可复现结果

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Training data supplier changes will destroy comparability

It is recommended to execute in the following order:

1. Check the actual version and source.
2. Use a test account, test directory, or test device.
3. Perform an undoable action.
4. Save the output, exit code, or screenshot.
5. Revert the changes and re-verify.

1

# 固定供应商、下载时间与字段定义

The completion standard is that “changes in training data suppliers will destroy comparability” and obtain verifiable results, rather than that the command happens to report no errors.

Prediction intervals and trading holding periods cannot be mixed.

1

# 分别记录 forecast horizon 与 holding period

When a stop signal appears, resume first and do not continue to the next step.

The backtest report must retain the failure interval

This section only deals with “the backtest report must retain the failure interval”. Save the current state first and then run the minimal test. Do not change multiple conditions at the same time.

1

# 报告最大回撤、换手率和分年度结果

For Kronos, this boundary must be combined when judging the results: Kronos is a financial K-line modeling project; the forecast output cannot be used as investment advice, and data leakage, re-righting methods, transaction costs and out-of-sample testing will significantly change the backtest results.

If the results are not repeatable, record the version, input, and logs, then narrow the scope of the test.

Final Check with Kronos

• The versions and sources used are documented.
• Test input and formal data have been isolated.
• Commands, configurations and actual output can correspond.
• The failed path was actively tested at least once.
• Keys, accounts and private data do not enter Git or logs.
• There are clear actions for upgrading, stopping and resuming.

If you still can’t explain how the results occurred, keep Kronos in a test environment without extending permissions or replacing existing stable processes.