quant-concept-drift

Quantitative Financeqlibrigorous codebase

Description

Concept Drift Adaptation in Stock Prediction on CSI300

Objective

Design and implement a stock prediction model that handles concept drift and temporal distribution shift in CSI300 stocks. Your code goes in custom_model.py. Three reference implementations (TRA, AdaRNN, LightGBM) are provided as read-only.

Evaluation

Signal quality: IC, ICIR, Rank IC. Portfolio (TopkDropout, top 50, drop 5): Annualized Return, Max Drawdown, Information Ratio. Automatic via qlib's workflow.

Evaluation uses three fixed temporal regimes on the same CSI300 universe:

csi300: long-horizon split ending in the 2017-2020 regime
csi300_shifted: shifted split with a 2016-2018 test regime
csi300_recent (hidden): the most recent 2019-2020 regime

This task is about temporal drift adaptation, not cross-universe transfer.

Workflow Configuration

workflow_config.yaml lines 13-26 and 32-45 are editable. This covers the model plus dataset adapter/processor configuration needed by methods like TRA. Instruments, date ranges, train/valid/test splits, and evaluation settings are fixed.

Code

custom_model.py

EditableRead-only

1# Custom stock prediction model for MLS-Bench (concept drift adaptation)
2#
3# EDITABLE section: CustomModel class with fit() and predict() methods.
4# FIXED sections: imports below.
5import numpy as np
6import pandas as pd
7import torch
8import torch.nn as nn
9import torch.nn.functional as F
10from qlib.model.base import Model
11from qlib.data.dataset import DatasetH
12from qlib.data.dataset.handler import DataHandlerLP
13
14DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
15

workflow_config.yaml

EditableRead-only

1# Qlib workflow configuration for CSI300 concept drift adaptation benchmark.
2# Used by run_workflow.py — default Alpha158/CSI300/DatasetH pipeline.
3# Alpha158: 158 engineered features per stock per day.
4
5qlib_init:
6  provider_uri: "~/.qlib/qlib_data/cn_data"
7  region: cn
8
9sys:
10  rel_path:
11    - "."           # So custom_model.py is importable via module_path
12
13task:
14  model:
15    class: CustomModel

Additional context files (read-only):

qlib/qlib/model/base.py

Results

Model	Type	ic csi300 shifted ↑	icir csi300 shifted ↑	rank ic csi300 shifted ↑	rank icir csi300 shifted ↑	annualized return csi300 shifted ↑	max drawdown csi300 shifted ↓	information ratio csi300 shifted ↑	ic csi300 recent ↑	icir csi300 recent ↑	rank ic csi300 recent ↑	rank icir csi300 recent ↑	annualized return csi300 recent ↑	max drawdown csi300 recent ↓	information ratio csi300 recent ↑	ic csi300 ↑	icir csi300 ↑	rank ic csi300 ↑	rank icir csi300 ↑	annualized return csi300 ↑	max drawdown csi300 ↓	information ratio csi300 ↑
adarnn	baseline	0.045	0.439	0.045	0.438	0.138	-0.081	1.815	0.017	0.154	0.028	0.253	0.021	-0.065	0.283	0.036	0.319	0.040	0.365	0.050	-0.104	0.673
adarnn	baseline	0.057	0.516	0.059	0.519	0.122	-0.081	1.482	0.017	0.169	0.028	0.274	0.000	-0.083	0.004	0.039	0.310	0.050	0.392	0.076	-0.087	1.195
adarnn	baseline	0.056	0.450	0.061	0.483	0.151	-0.148	1.691	0.026	0.207	0.040	0.329	0.084	-0.049	1.118	0.040	0.331	0.050	0.413	0.097	-0.082	1.399
lgbm	baseline	0.055	0.430	0.064	0.510	0.103	-0.079	1.299	0.025	0.185	0.044	0.322	0.081	-0.051	1.104	0.046	0.364	0.057	0.466	0.092	-0.058	1.347
lgbm	baseline	0.055	0.430	0.064	0.510	0.103	-0.079	1.299	0.025	0.185	0.044	0.322	0.081	-0.051	1.104	0.046	0.364	0.057	0.466	0.092	-0.058	1.347
lgbm	baseline	0.055	0.430	0.064	0.510	0.103	-0.079	1.299	0.025	0.185	0.044	0.322	0.081	-0.051	1.104	0.046	0.364	0.057	0.466	0.092	-0.058	1.347
lgbm	baseline	0.055	0.430	0.064	0.510	0.103	-0.079	1.299	0.025	0.185	0.044	0.322	0.081	-0.051	1.104	0.046	0.364	0.057	0.466	0.092	-0.058	1.347
tra	baseline	0.051	0.428	0.055	0.459	0.122	-0.084	1.587	0.024	0.192	0.040	0.330	0.033	-0.071	0.480	0.042	0.335	0.050	0.423	0.064	-0.100	1.008
tra	baseline	0.051	0.428	0.055	0.459	0.122	-0.084	1.587	0.024	0.192	0.040	0.330	0.033	-0.071	0.480	0.042	0.335	0.050	0.423	0.064	-0.100	1.008
tra	baseline	0.051	0.428	0.055	0.459	0.122	-0.084	1.587	0.024	0.192	0.040	0.330	0.033	-0.071	0.480	0.042	0.335	0.050	0.423	0.064	-0.100	1.008
anthropic/claude-opus-4.6	vanilla	0.054	0.408	0.065	0.500	0.157	-0.088	1.841	0.025	0.175	0.046	0.324	0.048	-0.058	0.655	0.043	0.323	0.056	0.426	0.102	-0.055	1.420
google/gemini-3.1-pro-preview	vanilla	0.047	0.326	0.063	0.427	0.149	-0.082	1.764	0.016	0.105	0.037	0.238	-0.016	-0.094	-0.219	0.041	0.292	0.055	0.384	0.008	-0.119	0.113
gpt-5.4-pro	vanilla	0.020	0.118	0.034	0.190	-0.050	-0.354	-0.457	0.019	0.131	0.036	0.265	0.031	-0.076	0.402	0.012	0.075	0.029	0.171	-0.086	-0.361	-0.823
anthropic/claude-opus-4.6	agent	0.054	0.400	0.066	0.497	0.189	-0.086	2.210	0.024	0.168	0.046	0.322	0.051	-0.052	0.699	0.043	0.319	0.055	0.422	0.060	-0.114	0.843
google/gemini-3.1-pro-preview	agent	0.047	0.326	0.063	0.427	0.149	-0.082	1.764	0.016	0.105	0.037	0.238	-0.016	-0.094	-0.219	0.041	0.292	0.055	0.384	0.008	-0.119	0.113
gpt-5.4-pro	agent	0.050	0.344	0.067	0.501	0.167	-0.040	2.153	0.020	0.148	0.044	0.313	-0.030	-0.121	-0.406	0.029	0.198	0.051	0.368	-0.011	-0.134	-0.149

Agent Conversations

anthropic/claude-opus-4.6

7 steps

google/gemini-3.1-pro-preview

6 steps

gpt-5.4-pro

6 steps