Qlib Workflow for Training and Backtesting an Alpha360 LSTM Model
Summary
This configuration specifies a Qlib research workflow for China A-share stocks in the CSI 300 universe. It uses Alpha360 features, robust feature normalization and missing-value filling, and cross-sectional rank normalization for labels. The prediction target is a forward close-to-close return, while the dataset is split into training, validation, and test periods. Sequences contain 60 steps, and the model uses an LSTM with attention alongside a separate recurrent component configured for three states.
The workflow also defines a top-k dropout portfolio strategy that holds 50 names and drops five, with a close-price backtest against the CSI 300 benchmark. Transaction costs, a minimum fee, and a price-limit threshold are specified. Signal analysis and portfolio analysis records are requested. This is an experiment recipe, not a report of findings: it supplies no measured returns, benchmark comparison, or evidence that the model generalizes. Results would depend on the data, implementation, and execution assumptions used to run it.
Key ideas
- The workflow configures an Alpha360 model for the CSI 300 universe using Qlib.
- It defines a forward close-to-close return target and separate training, validation, and test periods.
- The prediction model uses recurrent neural network components, including an LSTM with attention.
- A top-k dropout strategy and transaction-cost assumptions are specified for portfolio backtesting.
- The configuration describes an experiment setup but provides no performance results.
Tags
Full text
# workflow_config_tra_Alpha360.yaml
```yaml
qlib_init:
provider_uri: "~/.qlib/qlib_data/cn_data"
region: cn
market: &market csi300
benchmark: &benchmark SH000300
data_handler_config: &data_handler_config
start_time: 2008-01-01
end_time: 2020-08-01
fit_start_time: 2008-01-01
fit_end_time: 2014-12-31
instruments: *market
infer_processors:
- class: RobustZScoreNorm
kwargs:
fields_group: feature
clip_outlier: true
- class: Fillna
kwargs:
fields_group: feature
learn_processors:
- class: CSRankNorm
kwargs:
fields_group: label
label: ["Ref($close, -2) / Ref($close, -1) - 1"]
num_states: &num_states 3
memory_mode: &memory_mode sample
tra_config: &tra_config
num_states: *num_states
rnn_arch: LSTM
hidden_size: 32
num_layers: 1
dropout: 0.0
tau: 1.0
src_info: LR_TPE
model_config: &model_config
input_size: 6
hidden_size: 64
num_layers: 2
rnn_arch: LSTM
use_attn: True
dropout: 0.0
port_analysis_config: &port_analysis_config
strategy:
class: TopkDropoutStrategy
module_path: qlib.contrib.strategy
kwargs:
signal: <PRED>
topk: 50
n_drop: 5
backtest:
start_time: 2017-01-01
end_time: 2020-08-01
account: 100000000
benchmark: *benchmark
exchange_kwargs:
limit_threshold: 0.095
deal_price: close
open_cost: 0.0005
close_cost: 0.0015
min_cost: 5
task:
model:
class: TRAModel
module_path: qlib.contrib.model.pytorch_tra
kwargs:
tra_config: *tra_config
model_config: *model_config
model_type: RNN
lr: 1e-3
n_epochs: 100
max_steps_per_epoch:
early_stop: 20
logdir: output/Alpha360
seed: 0
lamb: 1.0
rho: 0.99
alpha: 0.5
transport_method: router
memory_mode: *memory_mode
eval_train: False
eval_test: True
pretrain: True
init_state:
freeze_model: False
freeze_predictors: False
dataset:
class: MTSDatasetH
module_path: qlib.contrib.data.dataset
kwargs:
handler:
class: Alpha360
module_path: qlib.contrib.data.handler
kwargs: *data_handler_config
segments:
train: [2008-01-01, 2014-12-31]
valid: [2015-01-01, 2016-12-31]
test: [2017-01-01, 2020-08-01]
seq_len: 60
input_size: 6
num_states: *num_states
batch_size: 1024
n_samples:
memory_mode: *memory_mode
drop_last: True
record:
- class: SignalRecord
module_path: qlib.workflow.record_temp
kwargs:
model: <MODEL>
dataset: <DATASET>
- class: SigAnaRecord
module_path: qlib.workflow.record_temp
kwargs:
ana_long_short: False
ann_scaler: 252
- class: PortAnaRecord
module_path: qlib.workflow.record_temp
kwargs:
config: *port_analysis_config
```Shown in full with attribution under the source's licence. Licence: MIT
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.