Usage

Command-line interface

timer-fit <working_directory> [options]

Option	Description
-v, --verbose	Enable verbose console output
-o, --outdir	Output directory name (default: out)

Examples

timer-fit examples/hip67522b
timer-fit examples/hip67522b -v
timer-fit examples/hip67522b -o model1

Python API

import yaml
from timer.fit import TransitFit

sys_params = yaml.load(open('sys.yaml'), Loader=yaml.FullLoader)
fit_params = yaml.load(open('fit.yaml'), Loader=yaml.FullLoader)

fit = TransitFit(sys_params, fit_params, wd='.')
fit.build_model()
fit.clip_outliers()
fit.sample()
fit.plot_corner()
fit.save_results()

Loading saved results

from timer.fit import TransitFit

fit = TransitFit.from_dir('examples/hip67522b')

Pipeline

The timer-fit CLI runs the following steps in order:

1. Load data -- read light curve files, bin, detrend
2. Build model -- construct PyMC model with priors, optimize for MAP solution
3. Clip outliers -- sigma-clip residuals (if clip: true in data config)
4. Re-fit -- rebuild model with outlier mask applied
5. Sample -- MCMC sampling with PyMC
6. Plot -- light curve fits, corner plots, trace plots, limb darkening
7. Save -- summary statistics, transit times, posterior samples, corrected light curves

Outputs

All outputs are saved to the out/ directory (or custom --outdir):

File	Description
fit.png	Light curve fit with residuals
corner.png	Corner plot of posterior distributions
trace.png	MCMC trace plot
summary.csv	Parameter summary statistics
tc.txt	Fitted transit center times
ic.txt	Information criteria (BIC, AIC, AICc), and for GP fits their effective degrees of freedom corrected counterparts
posterior_samples.csv.gz	Full posterior samples
*-cor.csv	Corrected (detrended) light curves
timer-fit.log	Full log file
fit.yaml, sys.yaml	Copies of input configuration
cache.json	Records which config and data each cached artifact was produced from

Reading ic.txt

The criteria are built from the maximum observed-data log likelihood over the posterior draws, so the penalty is nparams * log(ndata) and nothing else. In particular they do not include the prior terms: a criterion that did would charge each systematics column the log density of its own weight prior, about four times the intended penalty, and would move if that prior were widened.

Values written before this change are not comparable with values written after it. Recompute rather than mixing them.

For GP fits, edf is the number of degrees of freedom the GP smoother actually absorbs, typically far more than its two or three hyperparameters, and the *_edf rows repeat each criterion with that substituted. Treat nparams_edf as an upper bound: it does not subtract the overlap between the GP and the linear systematics model, so a fit with a wide design matrix is charged for some flexibility twice.

Reading *-cor.csv

yerr is the uncertainty the likelihood used, the photometric error and the fitted jitter added in quadrature, not the bare photometric error. Where a light curve was binned with the default kind='median', the binned error also carries the sqrt(pi/2) factor that the median of a sample costs over its mean.