Implementation:Shiyu coder Kronos TopkDropoutStrategy Usage

Field	Value
Implementation Name	TopkDropoutStrategy_Usage
Repository	Shiyu_coder_Kronos
Repository URL	https://github.com/shiyu-coder/Kronos
Type	External Tool Doc
Source	External -- qlib.contrib.strategy.TopkDropoutStrategy (not defined in this repo)
Used In	finetune/qlib_test.py:L120-125
Implements Principle	Principle:Shiyu_coder_Kronos_Topk_Dropout_Portfolio_Strategy
External Docs	https://qlib.readthedocs.io/
Last Updated	2026-02-09 14:00 GMT

Overview

TopkDropoutStrategy is an external trading strategy class from the Qlib quantitative investment framework. It is used in the Kronos backtesting pipeline to convert model prediction signals into portfolio allocation decisions. This class is not defined in the Kronos repository but is imported and configured within it.

Import

from qlib.contrib.strategy import TopkDropoutStrategy

API

strategy = TopkDropoutStrategy(
    topk: int,
    n_drop: int,
    hold_thresh: int,
    signal: pd.Series
) -> TopkDropoutStrategy

Parameters

Parameter	Type	Kronos Default	Description
topk	int	50	Number of stocks to hold in the portfolio at each rebalancing step
n_drop	int	5	Number of positions to randomly drop from the held pool at each rebalance
hold_thresh	int	5	Minimum holding period (in trading days) before a position is eligible for dropout
signal	pd.Series	(from model)	Prediction scores as a MultiIndex Series with levels (instrument, datetime)

Input: Signal Series

The signal parameter expects a pandas Series with a MultiIndex of (instrument, datetime) and float prediction scores as values. In the Kronos pipeline, this signal is generated from model inference:

# From finetune/qlib_test.py -- signal construction
pred_series = pred_df.stack()
pred_series.index.names = ['datetime', 'instrument']
pred_series = pred_series.swaplevel().sort_index()

# pred_series is now indexed as (instrument, datetime) with prediction scores

Output

The strategy object is passed to qlib.backtest.backtest() which uses it to generate portfolio allocation decisions at each time step. The backtest returns portfolio metrics including returns, costs, and benchmark comparisons.

Configuration in Kronos

The strategy parameters are configured via the Kronos Config class in finetune/config.py:

# From finetune/config.py
self.backtest_n_symbol_hold = 50   # Number of symbols to hold in the portfolio.
self.backtest_n_symbol_drop = 5    # Number of symbols to drop from the pool.
self.backtest_hold_thresh = 5      # Minimum holding period for a stock.

Usage in Kronos Backtesting

The complete usage pattern from finetune/qlib_test.py (L120-125):

from qlib.contrib.strategy import TopkDropoutStrategy
from qlib.backtest import backtest, executor

# Create strategy from prediction signals
strategy = TopkDropoutStrategy(
    topk=self.config.backtest_n_symbol_hold,
    n_drop=self.config.backtest_n_symbol_drop,
    hold_thresh=self.config.backtest_hold_thresh,
    signal=signal_series,
)

# Configure executor
executor_config = {
    "time_per_step": "day",
    "generate_portfolio_metrics": True,
    "delay_execution": True,
}

# Configure backtest
backtest_config = {
    "start_time": self.config.backtest_time_range[0],
    "end_time": self.config.backtest_time_range[1],
    "account": 100_000_000,
    "benchmark": self.config.backtest_benchmark,
    "exchange_kwargs": {
        "freq": "day",
        "limit_threshold": 0.095,
        "deal_price": "open",
        "open_cost": 0.001,
        "close_cost": 0.0015,
        "min_cost": 5,
    },
    "executor": executor.SimulatorExecutor(**executor_config),
}

# Run backtest
portfolio_metric_dict, _ = backtest(strategy=strategy, **backtest_config)

Signal Types

The Kronos inference pipeline generates multiple signal types from the same model predictions:

Signal Type	Computation	Description
last	`preds[:, -1, 3] - last_day_close`	Predicted close price change at the last prediction step
mean	`np.mean(preds[:, :, 3], axis=1) - last_day_close`	Average predicted close price change across all prediction steps
max	`np.max(preds[:, :, 3], axis=1) - last_day_close`	Maximum predicted close price change
min	`np.min(preds[:, :, 3], axis=1) - last_day_close`	Minimum predicted close price change

Each signal type can be independently backtested to evaluate which prediction aggregation method yields the best trading performance.

External Documentation

For full documentation on TopkDropoutStrategy and the Qlib backtesting framework, refer to:

Qlib documentation: https://qlib.readthedocs.io/
Qlib GitHub: https://github.com/microsoft/qlib

Environment & Heuristic Links

Environment:Shiyu_coder_Kronos_Qlib_Data_Environment

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment