CHANGES.md

v0.30 - development version

• Added documentation for model functions and sq.sample in README.
• Adds support for Python 3.13, drops support for Python 3.9 and 3.10.
• Uses prettier tqdm output that is now aware of Jupyter notebooks.
• bayes.update now supports lognorm and gamma distributions in addition to norm and beta.
• Allow zero_inflated to accept 0 or 1 as probabilities.
• [New feature] Added die(sides, explode_on=None) and coin() as distribution objects for dice and coins. These support all distribution operations like ~, @, +, -, etc. The die distribution also supports "exploding dice" mechanics via the explode_on parameter.
• [Breaking change] Removed roll_die and flip_coin functions. Use sq.die(sides) @ n or ~sq.coin() instead.

v0.29 - latest release

• Fixes a bug where max_gain and modeled_gain were incorrect in kelly output.
• Fixes a bug where error was not correctly passed from half_kelly, third_kelly, and quarter_kelly.
• Added invlognorm as a new distribution.
• Added bucket_percentages to more easily get the percentage of values within a bucket.
• Added third_kelly as an alias for kelly with deference = 0.66.
• Allows Bernoulli distributions to be defined with p=0 or p=1
• Added a Makefile to help simplify testing and linting workflows

v0.28

• [Breaking change] sq.pareto previously sampled from a Lomax distribution due to a complication with numpy. Now it properly samples from a Pareto distribution.
• [Breaking change] lclip / rclip have been removed from triangular distribution because that doesn't make sense.
• [Breaking change] You now can nest mixture and discrete distributions within mixture distributions.
• [Breaking change] sq.kelly now raises an error if you put in a price below the market price. You can pass error=False to disable this and return to the old behavior.
• Added pert distribution.
• Added sharpe_ratio to utilities.
• get_percentiles, get_log_percentiles, get_mean_and_ci, and get_median_and_ci now can all take an optional weights parameter to do a weighted version.

v0.27

• [Breaking change] This package now only supports Python 3.9 and higher.
• [Breaking change] get_percentiles and get_log_percentiles now always return a dictionary, even if there's only one element.
• [Breaking change] .type is now removed from distribution objects.
• You can now create correlated variables using sq.correlate.
• Added geometric distribution.
• Distribution objects now have the version of squigglepy they were created with, which can be accessed via obj._version. This should be helpful for debugging and noticing stale objects, especially when squigglepy distributions are stored in caches.
• Distributions can now be hashed with hash.
• Fixed a bug where tdist would not return multiple samples if defined with t alone.
• Package load time is now ~2x faster.
• Mixture sampling is now ~2x faster.
• Pandas and matplotlib as removed as required dependencies, but their related features are lazily enabled when the modules are available. These packages are still available for install as extras, installable with pip install squigglepy[plots] (for plotting-related functionality, matplotlib for now), pip install squigglepy[ecosystem] (for pandas, and in the future other related packages), or pip install squigglepy[all] (for all extras).
• Multicore distribution now does extra checks to avoid crashing from race conditions.
• Using black now for formatting.
• Switched from flake8 to ruff.

v0.26

• [Breaking change] lognorm can now be defined either referencing the mean and sd of the underlying normal distribution via norm_mean / norm_sd or via the mean and sd of the lognormal distribution itself via lognorm_mean / lognorm_sd. To further disambiguate, mean and sd are no longer variables that can be passed to lognorm.

v0.25

• Added plot as a method to more easily plot distributions.
• Added dist_log and dist_exp operators on distributions.
• Added growth_rate_to_doubling_time and doubling_time_to_growth_rate convenience functions. These take numbers, numpy arrays or distributions.
• Mixture distributions now print with weights in addition to distributions.
• Changes get_log_percentiles to report in scientific notation.
• bayes now supports separate arguments for memcache_load and memcache_save to better customize how memcache behavior works. memcache remains a parameter that sets both memcache_load and memcache_save to True.

v0.24

• Distributions can now be negated with - (e.g., -lognorm(0.1, 1)).
• Numpy ints and floats can now be used for determining the number of samples.
• Fixed some typos in the documentation.

v0.23

• Added pareto distribution.
• Added get_median_and_ci to return the median and a given confidence interval for data.
• discrete and mixture distributions now give more detail when printed.
• Fixed some typos in the documentation.

v0.22

• Added extremize to extremize predictions.
• Added normalize to normalize a list of numbers to sum to 1.
• Added get_mean_and_ci to return the mean and a given confidence interval for data.
• Added is_dist to determine if an object is a Squigglepy distribution.
• Added is_sampleable to determine if an object can be sampled using sample.
• Support for working within Pandas is now explicitly added. pandas has been added as a requirement.
• discrete sampling now will compress a large array if possible for more efficient sampling.
• clip, lclip, and rclip can now be used without needing distributions.
• Some functions (e.g, geomean) previously only supported lists, dictionaries, and numpy arrays. They have been expanded to support all iterables.
• dist_max and dist_min now support pipes (>>)
• get_percentiles now coerces output to integer if digits is less than or equal to 0, instead of just exactly 0.

v0.21

• Mixture sampling is now 4-23x faster.
• You can now get the version of squigglepy via sq.__version__.
• Fixes a bug where the tqdm was displayed with the incorrect count when collecting cores during a multicore sample.

v0.20

• Fixes how package dependencies are handled in setup.py an specifies Python >= 3.7 must be used. This should fix install errors.

v0.19

Bugfixes

• Fixes a bug where lclip and/or rclip on mixture distribution were not working correctly.
• Fixes a bug where dist_fn did not work with np.vectorize functions.
• Fixes a bug where in-memory caching was invoked for bayesnet when not desired.

Caching and Multicore

• [Breaking change] bayesnet caching is now based on binary files instead of pickle files (uses msgspec as the underlying library).
• [Breaking change] sample caching is now based on numpy files instead of pickle files.
• A cache can now be loaded via sample(load_cache=cachefile) or bayesnet(load_cache=cachefile), without needing to pass the distribution / function.
• bayesnet and sample now take an argument cores (default 1). If greater than 1, will run the calculations on multiple cores using the pathos package.

Other

• Functions that take weights now can instead take a parameter relative_weights where waits are automatically normalized to sum to 1 (instead of erroring, which is still the behavior if using weights).
• Verbose output for bayesnet and sample is now clearer (and slightly more verbose).

v0.18

• [Breaking change] The default t for t-distributions has changed from 1 to 20.
• sample results can now be cached in-memory using memcache=True. They can also be cached to a file -- use dump_cache_file to write the file and load_cache_file to load from the file.
• (Non-visible backend change) Weights that are set to 0 are now dropped entirely.

v0.17

• When verbose=True is used in sample, the progress bar now pops up in more relevant places and is much more likely to get triggered when relevant.
• discrete_sample and mixture_sample now can take verbose parameter.

v0.16

• zero_inflated can create an arbitrary zero-inflated distribution.
• Individual sampling functions (normal_sample, lognormal_sample, etc.) can now take an argument samples to generate multiple samples.
• A large speedup has been achieved to sampling from the same distribution multiple times.
• requirements.txt has been updated.

v0.15

• [Breaking change] bayesnet function now refers to parameter memcache where previously this parameter was called cache.
• [Breaking change] If get_percentiles or get_log_percentiles is called with just one elemement for percentiles, it will return that value instead of a dict.
• Fixed a bug where get_percentiles would not round correctly.
• bayesnet results can now be cached to a file. Use dump_cache_file to write the file and load_cache_file to load from the file.
• discrete now works with numpy arrays in addition to lists.
• Added one_in as a shorthand to convert percentages into "1 in X" notation.
• Distributions can now be compared with == and !=.

v0.14

• Nested sampling now works as intended.
• You can now use >> for pipes for distributions. For example, sq.norm(1, 2) >> dist_ceil
• Distributions can now be compared with >, <, >=, and <=.
• dist_max can be used to get the maximum value between two distributions. This family of functions are not evaluated until the distribution is sampled and they work with pipes.
• dist_min can be used to get the minimum value between two distributions.
• dist_round can be used to round the final output of a distribution. This makes the distribution discrete.
• dist_ceil can be used to ceiling round the final output of a distribution. This makes the distribution discrete.
• dist_floor can be used to floor round the final output of a distribution. This makes the distribution discrete.
• lclip can be used to clip a distribution to a lower bound. This is the same functionality that is available within the distribution and the sample method.
• rclip can be used to clip a distribution to an upper bound. This is the same functionality that is available within the distribution and the sample method.
• clip can be used to clip a distribution to both an upper bound and a lower bound. This is the same functionality that is available within the distribution and the sample method.
• sample can now be used directly on numbers. This makes const functionally obsolete, but const is maintained for backwards compatibility and in case it is useful.
• sample(None) now returns None instead of an error.

v0.13

• Sample shorthand notation can go in either order. That is, 100 @ sq.norm(1, 2) now works and is the same as sq.norm(1, 2) @ 100, which is the same as sq.sample(sq.norm(1, 2), n=100).

v0.12

• Distributions now implement math directly. That is, you can do things like sq.norm(2, 3) + sq.norm(4, 5), whereas previously this would not work. Thanks to Dawn Drescher for helping me implement this.
• ~sq.norm(1, 2) is now a shorthand for sq.sample(sq.norm(1, 2)). Thanks to Dawn Drescher for helping me implement this shorthand.
• sq.norm(1, 2) @ 100 is now a shorthand for sq.sample(sq.norm(1, 2), n=100)

v0.11

Distributions

• [Breaking change] tdist and log_tdist have been modified to better approximate the desired credible intervals.
• tdist now can be defined by just t, producing a classic t-distribution.
• tdist now has a default value for t: 1.
• Added chisquare distribution.
• lognormal now returns an error if it is defined with a zero or negative value.

Other

• All functions now have docstrings.
• Added kelly to calculate Kelly criterion for bet sizing with probabilities.
• Added full_kelly, half_kelly, quarter_kelly as helpful aliases.

v0.10

• [Breaking change] credibility is now defined using a number out of 100 (e.g., credibility=80 to define an 80% CI) rather than a decimal out of 1 (e.g., credibility=0.8 to define an 80% CI).
• Distribution objects now print their parameters.

v0.9

• goemean and geomean_odds now can take the nested-list-based and dictionary-based formats for passing weights.

v0.8

Bayesian library updates

• [Breaking change] bayes.update now updates normal distributions from the distribution rather than from samples.
• [Breaking change] bayes.update no longer takes a type parameter but can now infer the type from the passed distribution.
• [Breaking change] Corrected a bug in how bayes.update implemented evidence_weight when updating normal distributions.

Non-visible backend changes

• Distributions are now implemented as classes (rather than lists).

v0.7

Bugfixes

• Fixes an issue with sampling from the bernoulli distribution.
• Fixes a bug with the implementation of lclip and rclip.

New distributions

• Adds discrete to calculate a discrete distribution. Example: discrete({'A': 0.3, 'B': 0.3, 'C': 0.4}) will return A 30% of the time, B 30% of the time, and C 40% of the time.
• Adds poisson(lam) to calculate a poisson distribution.
• Adds gamma(size, scale) to calculate a gamma distribution.

Bayesian library updates

• Adds bayes.bayesnet to do bayesian inferece (see README).
• bayes.update now can take an evidence_weight parameter. Typically this would be equal to the number of samples.
• [Breaking change] bayes.bayes has been renamed bayes.simple_bayes.

Other

• [Breaking change] credibility, which defines the size of the interval (e.g., credibility=0.8 for an 80% CI), is now a property of the distribution rather than the sampler. That is, you should now call sample(norm(1, 3, credibility=0.8)) whereas previously it was sample(norm(1, 3), credibility=0.8). This will allow mixing of distributions with different credible ranges.
• [Breaking change] Numbers have been changed from functions to global variables. Use thousand or K instead of thousand() (old/deprecated).
• sample now has a nice progress reporter if verbose=True.
• The exponential distribution now implements lclip and rclip.
• The mixture distribution can infer equal weights if no weights are given.
• The mixture distribution can infer the last weight if the last weight is not given.
• geomean and geomean_odds can infer the last weight if the last weight is not given.
• You can use flip_coin and roll_die(sides) to flip a coin or roll a die.
• event_happens and event are aliases for event_occurs.
• get_percentiles will now cast output to int if digits=0.
• get_log_percentiles now has a default value for percentiles.
• You can now set the seed for the RNG using sq.set_seed.

Non-visible backend changes

• Now has tests via pytest.
• The random numbers now come from a numpy generator as opposed to the previous deprecated np.random methods.
• The sample module (containing the sample function) has been renamed samplers.

v0.6

New distributions

• Add triangular(left, mode, right) to calculate a triangular distribution.
• Add binomial(n, p) to calculate a binomial distribution.
• Add beta(a, b) to calculate a beta distribution.
• Add bernoulli(p) to calculate a bernoulli distribution.
• Add exponential(scale) to calculate an exponential distribution.

New Bayesian library

• Add bayes.update to get a posterior distribution from a prior distribution and an evidence distribution.
• Add bayes.average to average distributions (via a mixture).

New utility functions

• Add laplace to calculate Laplace's Law of Succession. If s and n are passed, it will calculate (s+1)/(n+2). If s, time_passed, and time_remaining are passed, it will use the time invariant version. Use time_fixed=True for fixed time periods and time_fixed=False (default) otherwise.
• Add geomean to calculate the geometric mean.
• Add p_to_odds to convert probability to odds. Also odds_to_p to convert odds to probability.
• Add geomean_odds to calculate the geometric mean of odds, converted to and from probabilities.

Other

• If a distribution is defined with sd but not mean, mean will be inferred to be 0.
• sample can now take lclip and rclip directly, in addition to defining lclip and rclip on the distribution itself. If both are defined, the most restrictive of the two bounds will be used.

v0.5

• Fix critical bug to tdist and log_tdist introduced in v0.3.

v0.4

• Fix critical bug introduced in v0.3.

v0.3

• Be able to define distributions using mean and sd instead of defining the interval.

v0.2

• [Breaking change] Change distributed_log to mixture (to follow Squiggle) and allow it to implement any sub-distribution.
• [Breaking change] Changed library to single import.
• [Breaking change] Remove weighted_log as a distribution.

v0.1

• Initial library