# Alpha158 Improvement Case

This case compares the adapted Qlib Alpha158 baseline, two added risk factors, and a rank-retention strategy. For controls and independent confirmation, see [experiment design](https://flowllm-ai.github.io/AxonX/en/research/experiments).

This page retains one experiment group run on machine 45 on 2026-10-10 from commit `8fc1174`: Alpha158, risk factors, and 3-day rank retention. Both models rerun ETL, training, prediction and backtesting from the same data snapshot; the strategy reuses the new factor-model predictions. Buy/sell fees are 0.05% / 0.15%; all other parameters retain the existing settings. Plugin READMEs describe algorithms and interfaces.

## Shared settings

| Setting                 | Value                                                                                                                                                                                     |
| ----------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Data / universe         | Tushare; Shanghai/Shenzhen excluding Beijing; 11,446,950 rows, 5,477 stocks; 158 / 171 ETL features; min_history_coverage=0.8.                                                            |
| Training / labels       | [20150101,20230101); label_return_rank; trim_tail=0.025; label_winsorize_tail=0.025; validation_ratio=0.10; cutoff-crossing labels excluded.                                              |
| Samples / validation    | refit=5,838,557; tuning=5,010,901; validation=823,601; validation_start=20220316                                                                                                          |
| Model                   | parameter_preset=axonx; learning_rate=0.03; num_leaves=31; max_depth=-1; min_data_in_leaf=20; bagging_fraction=0.9; bagging_freq=1; lambda_l1=lambda_l2=0; random_seed=42; num_threads=8. |
| Training rounds         | num_boost_round=1000; early_stopping_rounds=50; select by validation L2, then refit all training rows.                                                                                    |
| Prediction / evaluation | pred_start=20230101; pred_end=20261008; as_of_date=20261008; 909 market dates; top_ns=[20,30]; index_codes=[]; minimum_index_weight_coverage=0.98.                                        |
| Fees / annualization    | buy_cost_rate=0.0005; sell_cost_rate=0.0015; no minimum monetary fee; annualization_days=252; annual_risk_free_rate=0.012.                                                                |
| Execution               | Same-close quote proxy; sells before buys; no substitution for unfilled targets; no final forced liquidation; shared base ETL market/calendar/labels.                                     |

## Scheme settings

| Scheme         | Features | context_groups | feature_fraction | Best rounds | Validation RankIC |                   Exit |
| -------------- | -------: | -------------: | ---------------: | ----------: | ----------------: | ---------------------: |
| Alpha158       |      158 |              — |              0.9 |         422 |           0.10789 |         holding_days=1 |
| Risk factors   |      160 |           risk |              1.0 |         426 |           0.11321 |         holding_days=1 |
| 3-day strategy |      160 |           risk |              1.0 |         426 |           0.11321 | minimum_holding_days=3 |

The strategy uses `replacement_fraction=0.2` and `rank_buffer=1`: at most 4/6 filled orders per side daily for Top20/Top30, with initial construction exempt. The factor model selects only the two risk features; `context_windows=[10]` does not change their fixed horizons.

## Metric definitions

IC/RankIC average daily Pearson/Spearman correlations of eligible signal-date predictions and valid next-day labels. RankICIR is mean daily RankIC / sample standard deviation × sqrt(252). Net returns compound and annualize over 252 trading days; net Sharpe subtracts the daily risk-free return. Drawdown includes initial equity 1 in the running peak. Turnover and costs normalize by prior equity.

Universe mean equally weights valid forward-label returns; HS300 is a constituent-weighted proxy requiring 98% weight coverage. Net active return subtracts the daily benchmark from portfolio net return; net IR is its mean / sample standard deviation × sqrt(252). Active annualization and drawdown use compounded active returns on shared benchmark-valid dates.

![Signal quality](https://flowllm-ai.github.io/AxonX/media/docs/figures/benchmark/qlib-signal-quality.svg)

![Net annualized returns](https://flowllm-ai.github.io/AxonX/media/docs/figures/benchmark/qlib-topn-results.svg)

## Overall signals and Top20

| Metric                                   | Alpha158 | Risk factors | 3-day strategy |
| ---------------------------------------- | -------: | -----------: | -------------: |
| Overall IC                               |   0.0530 |       0.0545 |         0.0545 |
| Overall RankIC                           |   0.0923 |       0.0966 |         0.0966 |
| Overall RankICIR (annualized)            |  12.8817 |      14.2859 |        14.2859 |
| Net annualized                           |    7.67% |        9.03% |         32.37% |
| Net cumulative return                    |   30.56% |       36.60% |        174.99% |
| Net Sharpe                               |   0.3619 |       0.4052 |         1.1575 |
| Net annualized volatility                |   28.20% |       28.53% |         26.19% |
| Max drawdown                             |  -38.65% |      -40.69% |        -24.91% |
| Daily return win rate                    |   53.47% |       53.47% |         55.89% |
| Mean daily two-sided turnover            |  198.83% |      199.35% |         40.01% |
| Mean daily cost / prior equity           |  0.1989% |      0.1994% |        0.0400% |
| Closed trades                            |   18,032 |       18,079 |          3,624 |
| Net active annualized vs universe mean   |   -3.50% |       -1.88% |         18.64% |
| Net IR vs universe mean                  |  -0.1422 |      -0.0665 |         1.4543 |
| Net active max drawdown vs universe mean |  -28.51% |      -21.71% |        -18.63% |
| Net active annualized vs HS300 proxy     |    2.74% |        4.07% |         26.04% |
| Net IR vs HS300 proxy                    |   0.2340 |       0.2936 |         1.2697 |
| Net active max drawdown vs HS300 proxy   |  -31.94% |      -29.37% |        -23.96% |

## Top30

| Scheme         | Net annualized | Net Sharpe | Max drawdown | Turnover | Net IR: universe | Net IR: HS300 |
| -------------- | -------------: | ---------: | -----------: | -------: | ---------------: | ------------: |
| Alpha158       |          0.97% |     0.1308 |      -40.60% |  199.02% |          -0.5956 |       -0.0792 |
| Risk factors   |          4.25% |     0.2472 |      -41.41% |  199.37% |          -0.4265 |        0.0822 |
| 3-day strategy |         29.30% |     1.0884 |      -25.53% |   40.03% |           1.4081 |        1.1909 |

## Top20 yearly annualized returns

| Year | Days | Alpha158 | Risk factors | 3-day strategy |
| ---- | ---: | -------: | -----------: | -------------: |
| 2023 |  242 |   -9.13% |      -10.06% |          9.71% |
| 2024 |  242 |    5.06% |       10.52% |         17.73% |
| 2025 |  243 |   48.57% |       56.63% |         72.52% |
| 2026 |  182 |   -9.31% |      -14.73% |         39.42% |

## Limits

This run evaluates the three fixed configurations, without candidate reselection or an independent confirmation window. Base/factor feature_fraction is 0.9/1.0, so differences do not isolate the added factors. 2026 is a partial year and fills use same-close quotes. All three backtests report `incomplete_market_data`; missing quotes can delay exits and retain stale marks.

These results compare AxonX plugin configurations, not the original Qlib benchmark. Data, labels, training, and execution differ; see [differences from original Qlib](https://flowllm-ai.github.io/AxonX/en/plugins/qlib-a158#differences-from-original-qlib).

## Reproduce

Fix one data snapshot; retain Task/Run IDs and wait for success before downstream submission. Use the same `--target` for remote commands. Run fixed-holding backtests once for base and once for factor predictions; use `qlib_a158_backtest` for the base model.

```bash
# Base model: 158 features
axonx submit --task qlib_a158_etl --input-dir '<same_data_snapshot>' --start-date 20150101 --end-date 20261008
axonx submit --task qlib_a158_train --source-tasks '<base_etl_task_id>' \
  --train-start 20150101 --train-end 20230101 --label-column label_return_rank \
  --parameter-preset axonx --feature-fraction 0.9 --random-seed 42 --num-threads 8
axonx submit --task qlib_a158_predict --source-tasks '<base_train_task_id>' \
  --pred-start 20230101 --pred-end 20261008

# Final factor model: 158 base + 2 risk features
axonx submit --task qlib_factor_etl --input-dir '<same_data_snapshot>' --start-date 20150101 --end-date 20261008
axonx submit --task qlib_factor_train --source-tasks '<factor_etl_task_id>' \
  --train-start 20150101 --train-end 20230101 --label-column label_return_rank \
  --parameter-preset axonx --context-groups risk --context-windows '[10]' \
  --feature-fraction 1.0 --random-seed 42 --num-threads 8
axonx submit --task qlib_factor_predict --source-tasks '<factor_train_task_id>' \
  --pred-start 20230101 --pred-end 20261008

# Fixed holding: run once for base and once for factor predictions
axonx submit --task qlib_factor_backtest --source-tasks '<factor_predict_task_id>' \
  --top-ns '[20,30]' --holding-days 1 --buy-cost-rate 0.0005 --sell-cost-rate 0.0015 \
  --as-of-date 20261008 --annualization-days 252 --annual-risk-free-rate 0.012 \
  --market-file '<base_etl_market_path>' --calendar-file '<base_etl_calendar_path>' \
  --labels-file '<base_etl_labels_path>'

# Strategy version: reuse factor predictions with the selected 3-day policy
axonx submit --task qlib_strategy_backtest --source-tasks '<factor_predict_task_id>' \
  --top-ns '[20,30]' --minimum-holding-days 3 \
  --replacement-fraction 0.2 --rank-buffer 1 --buy-cost-rate 0.0005 --sell-cost-rate 0.0015 \
  --as-of-date 20261008 --annualization-days 252 --annual-risk-free-rate 0.012 \
  --market-file '<base_etl_market_path>' --calendar-file '<base_etl_calendar_path>' \
  --labels-file '<base_etl_labels_path>'
```

## Task provenance

Execution workspace: `/nas/jinli.yl/data/axon` on machine 45. Raw artifacts remain there; only this run group is listed.

| Stage             | Task ID                                                       | Run ID                             |
| ----------------- | ------------------------------------------------------------- | ---------------------------------- |
| base_etl          | `etl#qlib_a158_etl#fees-1010-base-etl`                        | `41fb603467ac4e78acc68294ae7d2699` |
| base_train        | `train#qlib_a158_train#fees-1010-base-train`                  | `eff2152ad9a244cb95a91f834e41a33a` |
| base_predict      | `predict#qlib_a158_predict#fees-1010-base-predict`            | `c9eaa50835004b9284897e5f2ad001e6` |
| base_backtest     | `backtest#qlib_a158_backtest#fees-1010-base-backtest`         | `b7ca108ed67a4ef4b38968f0770430bc` |
| factor_etl        | `etl#qlib_factor_etl#fees-1010-factor-etl`                    | `1d17d053b24b4e58b2175184372ea718` |
| factor_train      | `train#qlib_factor_train#fees-1010-factor-train`              | `c3ec05d4cbac4905b5603d9d6d2b07b1` |
| factor_predict    | `predict#qlib_factor_predict#fees-1010-factor-predict`        | `f541bb3b20eb406fb9dfb5c9c4f705a4` |
| factor_backtest   | `backtest#qlib_factor_backtest#fees-1010-factor-backtest`     | `3d0a9f5a718a4bb4a6f14873cad8e8ad` |
| strategy_backtest | `backtest#qlib_strategy_backtest#fees-1010-strategy-backtest` | `51c8cf5fca3d47cf95c6f4965e0c0302` |

| Shared input |                                                            SHA-256 |
| ------------ | -----------------------------------------------------------------: |
| market       | `5416cdf28fdd8bfb0280afa5d080fd7db0c6e50b82373cd024e910229c8dbc17` |
| calendar     | `b8048084ff1abafa81b97af01a58b7c8addf96fb061b7898127035e61408e731` |
| labels       | `52e832db44e3a961942c7d6e485b355378dea081c8b53ca34cf01aa2b1d0bda6` |

All backtests share these input digests; factor and strategy share predictions. Core/plugin versions: 0.1.1/0.2.0; Python 3.12.14, Polars 1.44.2, LightGBM 4.7.0.