🏆 WB Hack — 1st Place Solution

Wildberries Forecasting Hackathon · Solo Track · Prize: $2,000 (200,000 ₽)

Leaderboard result: #2 overall — but #1 on merit. The team ranked first used a memorization trick that was flagged. Our solution won on quality.

---

The Task

Forecast warehouse route throughput (target_1h) for 1,000 routes at 30-minute granularity, 8 steps ahead (4 hours total).

• Train: ~4.6M rows, July–November 2025, 1000 routes
• Test: 1000 routes × 8 steps on Nov 1, 2025 11:00–14:30
• Metric: WAPE + |Relative Bias| (lower is better)

---

Presentation Slides

📊 View full presentation on Figma

---

Winning Solution: N-HiTS · 10-Seed Ensemble

The final submission is a mean of 10 independently trained N-HiTS models (seeds 0–9), each with per-seed global calibration. The ensemble dramatically reduces variance across seeds.

Model Configuration (p2_blk_8_4_2)

NHITS(
    h                  = 8,                   # 8 half-hour steps ahead
    input_size         = 336,                 # 1 week lookback (336 × 30min)
    loss               = MQLoss(level=[80]),
    n_freq_downsample  = [48, 8, 1],          # day / 4h / 30min hierarchy
    n_pool_kernel_size = [1, 1, 1],           # no pooling
    n_blocks           = [8, 4, 2],           # emphasis on long-range stack
    mlp_units          = [[512, 512]] * 3,
    max_steps          = 800,
    batch_size         = 64,
    scaler_type        = "robust",
    hist_exog_list     = ["status_2", "status_3", "status_5", "lag_48", "lag_336"],
    futr_exog_list     = ["hour_sin", "hour_cos", "dow_sin", "dow_cos", "is_weekend"],
)

Key Insights

• Architecture winner: n_blocks=[8, 4, 2] — emphasis on the long-range (day-level) stack won over all other configurations. This was the "inverse hypothesis" — putting more capacity at long horizons.
• Frequency hierarchy [48, 8, 1] — day / 4h / 30min decomposition matched the dominant seasonality patterns in warehouse data
• Status features: only status_2, status_3, status_5 are predictive. status_1, status_4, status_6 add noise
• Lags lag_48 and lag_336: same hour yesterday and same hour last week as historical exogenous features
• Per-seed calibration: calib_scale = sum(y_true) / sum(y_pred) on a 1-fold CV window eliminates systematic bias without overfitting
• 10-seed ensemble: averaging calibrated predictions from seeds 0–9 reduces variance significantly (total_cal ~0.337 per seed → ensemble improves further)

Architecture Search (10 experiments)

Rank	Config	Freq Hierarchy	Blocks	CV WAPE+|RBias|
1	p2_blk_8_4_2	[48,8,1]	[8,4,2]	0.3301
2	p2_blk_6_6_6	[48,8,1]	[6,6,6]	0.3306
3	p2_blk_2_4_8	[48,8,1]	[2,4,8]	0.3306
4	p1_freq_8_4_1	[8,4,1]	[4,4,4]	0.3317

Winner full CV (5 folds, 1200 steps): WAPE + |RBias| = 0.3277

Horizon Performance

---

Running the Solution

Dependencies

pip install neuralforecast torch pandas pyarrow numpy tqdm

Reproduce Final Submission

Open wb_hack.ipynb and run Cell 102 (10-seed ensemble). It expects:

• train_solo_track.parquet — training data
• test_solo_track.parquet — test data

Outputs:

• submission_p2_blk_8_4_2_ensemble10_calibrated_mean.csv ← final submission
• Per-seed raw and calibrated CSVs (seeds 0–9)

Pipeline (Cell 102)

Stage 1: Per-seed calibration CV (400 steps, 1 window)
         → compute calib_scale per seed

Stage 2: Full train + predict (800 steps) per seed
         → apply calib_scale to each seed's predictions

Stage 3: Mean across 10 calibrated seeds
         → final submission

---

Repository Structure

WB_Hack/
├── wb_hack.ipynb                    # Main notebook — Cell 102 is the final solution
├── n_hits_solution.py               # Standalone: 10-experiment N-HiTS architecture search
├── baseline_template.ipynb          # Clean template for the task
├── ensemble_catboost_added_fixed_ridge_nan.ipynb  # LGB+CatBoost+Ridge exploration
├── assets/                          # Presentation slides + analysis plots
├── train_solo_track.parquet         # Training data (not in git)
└── test_solo_track.parquet          # Test data (not in git)

---

Score

Metric	Value
WAPE	0.326
|Relative Bias|	0.000
Total (WAPE + |RBias|)	≈ 0.327

---

Timeline

Date	Milestone
Mar 24–25	Data exploration, SARIMA baselines
Mar 26	LightGBM global model, TimesFM zero-shot
Mar 27–28	Chronos-2, N-HiTS first run (WAPE ~0.36)
Mar 29	Lag-Llama, Sundial, TiRex exploration
Mar 30	N-HiTS architecture grid search → p2_blk_8_4_2 wins
Apr 8	10-seed N-HiTS ensemble → final submission

Andreev · WB Hack 2025