ACEFormer Forecasting

Transformer with Empirical Mode Decomposition (EMD) preprocessing and ProbabilityAttention for stock price forecasting.

Paper: An End-to-End Structure with Novel Position Mechanism and Improved EMD for Stock Forecasting
Source: DurandalLee/ACEFormer

---

Overview

Property	Value
Model Name	aceformer-forecasting
Category	Time Series
Framework	PyTorch
Trainable	Yes (train from scratch)
Pretrained	No — requires user data
GPU Required	Optional (faster with GPU)
VRAM	~512 MB
Parameters	~100K
Inference Engine	Specialist (forecasting)

---

Capabilities

• Stock price prediction using close prices (and optionally volume)
• Custom EMD (Empirical Mode Decomposition) signal preprocessing
• ProbSparse attention (Informer-style) with distillation layers
• Temporal position parameters for time-aware learning
• Lightweight model suitable for CPU inference
• Rolling (autoregressive) prediction for horizons longer than predict_size

---

Training

ACEFormer trains from scratch on user-provided stock data. There are no pretrained weights — the model learns patterns specific to your dataset.

Input Format

CSV file with at least a close price column. Optionally include volume for better performance.

Column	Required	Description
close	Yes	Closing price
volume	Optional	Trading volume
date	Optional	Date column (not used by model, but helpful for reference)

Minimum dataset size: 50 rows (200+ recommended for good results)

Hyperparameters

Parameter	Type	Default	Range	Description
unit_size	int	30	10-100	Input window size (trading days)
predict_size	int	5	1-30	Prediction horizon (trading days)
embed_dim	int	64	32-256	Embedding dimension
forward_dim	int	256	64-1024	Feed-forward hidden dimension
dis_layer	int	3	1-5	Number of distillation layers
attn_layer	int	2	1-4	Number of full attention layers
dropout	float	0.1	0.0-0.5	Dropout rate
learning_rate	float	0.001	1e-4 to 0.01	Learning rate
iterations	int	2000	100-10000	Training iterations
batch_size	int	64	8-256	Batch size
target_column	string	close	—	Target column name in CSV

Training Example

Python SDK:

from colabhive import ColabHive

client = ColabHive(api_key="...", account_id="...")

# Upload stock data
dataset = client.datasets.upload(name="AAPL-daily", file="./stock_prices.csv")

# Train ACEFormer
job = client.training.create(
    model="aceformer-forecasting",
    dataset_id=dataset.id,
    hyperparameters={
        "unit_size": 30,
        "predict_size": 5,
        "iterations": 2000,
        "target_column": "close"
    }
)

# Wait for completion
job.wait()
print(job.get_metrics())   # {"eval_mse": ..., "eval_mae": ..., "eval_rmse": ..., "eval_r2": ...}

cURL:

curl -X POST https://api.colabhive.com/api/builder/v1/training/runs \
  -H "X-API-Key: $KEY" \
  -H "X-Account-ID: $ACCT" \
  -H "Content-Type: application/json" \
  -d '{
    "job_name": "aceformer-aapl",
    "model_config_id": "ACEFORMER_CONFIG_UUID",
    "dataset_id": "DATASET_UUID",
    "hyperparameters": {
      "unit_size": 30,
      "predict_size": 5,
      "iterations": 2000,
      "target_column": "close"
    }
  }'

Training Output

File	Description
model.pt	PyTorch state_dict (model weights)
config.json	Hyperparameters + preprocessing state (model_type: "aceformer")
summary.json	Training metrics (MSE, MAE, RMSE, R2)

Metrics

Metric	Description
train_loss	Training MSE loss
eval_mse	Validation Mean Squared Error
eval_mae	Validation Mean Absolute Error
eval_rmse	Validation Root Mean Squared Error
eval_r2	Validation R-squared

---

Inference

After training, register the model for inference and send time series data.

Register for Inference

curl -X POST https://api.colabhive.com/api/builder/v1/training/runs/RUN_ID/register-for-inference \
  -H "X-API-Key: $KEY" -H "X-Account-ID: $ACCT" \
  -H "Content-Type: application/json" \
  -d '{"name": "aapl-predictor", "description": "AAPL stock price forecaster", "visibility": "account"}'

Run Predictions

Univariate (close prices only):

curl -X POST https://api.colabhive.com/api/builder/v1/endpoints/ENDPOINT_ID/infer \
  -H "X-API-Key: $KEY" -H "X-Account-ID: $ACCT" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "values": [150.2, 151.0, 149.8, 150.5, 152.0, 151.3, 153.1, 152.7, 154.0, 153.5,
                 154.2, 153.8, 155.0, 154.5, 155.8, 155.2, 156.0, 155.5, 156.8, 156.2,
                 157.0, 156.5, 157.8, 157.2, 158.0, 157.5, 158.8, 158.2, 159.0, 158.5],
      "horizon": 5
    }
  }'

Multivariate (close + volume):

curl -X POST .../infer \
  -d '{
    "input": {
      "series": {
        "close": [150.2, 151.0, ...],
        "volume": [1200000, 1150000, ...]
      },
      "horizon": 5
    }
  }'

Response

{
  "result": {
    "forecasts": [159.3, 159.8, 160.1, 159.5, 160.4],
    "model_type": "ACEFormer",
    "horizon": 5,
    "inference_time_ms": 28.5
  }
}

---

Architecture

ACEFormer combines three key innovations:

1. Empirical Mode Decomposition (EMD): Custom preprocessing that extracts mid-curve signal from raw price data, reducing noise while preserving trend
2. ProbSparse Attention: Informer-style attention that selects top-k most informative queries, reducing O(n^2) to O(n log n)
3. Distillation Layers: Progressive sequence reduction via MaxPool1d + temporal parameters

Input (batch, 30, features)
  -> ExpandEmbedding (Conv1d)
  -> + PositionalEmbedding (sinusoidal)
  -> 3x [ProbSparse CrossAttention + Distilling + Temporal Param]
  -> 2x [Full CrossAttention]
  -> Linear -> (batch, reduced_seq, 1)

---

Tips

• Minimum data: 200+ rows for meaningful results. More data = better generalization.
• Horizon vs predict_size: If you request horizon > predict_size, the model uses autoregressive rolling prediction. This works but accuracy degrades for longer horizons.
• Feature selection: Including volume alongside close price typically improves performance.
• Training iterations: Start with 500-1000 for quick experiments, use 2000-5000 for production models.
• Unit size: 30 (one trading month) is the default and works well for daily data. Use 5-10 for intraday data.