Getting Started¶

Installation¶

pip install pycupfm

Requirements¶

Python ≥ 3.9
NumPy ≥ 1.21
SciPy ≥ 1.7
Pandas ≥ 1.3
Matplotlib ≥ 3.5
Seaborn ≥ 0.12

Your First Model¶

from pycupfm import CupFM
from pycupfm.datasets import load_grunfeld

# 1. Load data
df = load_grunfeld()

# 2. Create model
model = CupFM(
    n_factors=1,        # number of common factors
    bandwidth=3,        # Bartlett kernel bandwidth
    kernel='bartlett',  # kernel type
    max_iter=25,        # max iterations
)

# 3. Fit all 5 estimators
results = model.fit(
    y=df['linvest'],
    X=df[['lmvalue', 'lkstock']],
    panel_id=df['firm'],
    time_id=df['year'],
    var_names=['lmvalue', 'lkstock'],
    dep_var='linvest'
)

# 4. View results
results.summary()

# 5. Generate plots
model.plot(kind='all')

# 6. Export
results.to_latex()
results.to_excel('results.xlsx')

Data Format¶

PyCupFM expects long-format panel data:

panel_id	time_id	y	x1	x2
1	1	5.2	3.1	2.4
1	2	5.8	3.5	2.6
...	...	...	...	...
N	T	4.1	2.8	1.9

Requirements

Balanced panel: every unit must have exactly T observations
I(1) variables: all variables should be integrated of order 1
No missing values: remove or impute NaN before fitting

Automatic Factor Selection¶

# Let Bai-Ng (2002) IC choose the number of factors
model = CupFM(n_factors='auto', auto_rmax=8)
results = model.fit(y, X, panel_id, time_id)
print(f'Selected r = {results.r}')

Using Your Own Data¶

import pandas as pd
from pycupfm import CupFM

df = pd.read_csv('my_panel_data.csv')

model = CupFM(n_factors=2, bandwidth=5)
results = model.fit(
    y=df['gdp'],
    X=df[['investment', 'exports']],
    panel_id=df['country'],
    time_id=df['year']
)
results.summary()