Shapley-based decomposition to anatomize the of out-of-sample accuracy of time-series forecasting models
Project description
Introduction
This package implements the $\text{oShapley-VI}_p$ (out-of-sample Shapley-based variable importance), $\text{PBSV}_p$ (performance-based Shapley value), MAS (model accordance score) and MAS hypothesis testing, proposed in the "The Anatomy of Out-of-Sample Forecasting Accuracy" paper by Daniel Borup, Philippe Goulet Coulombe, David E. Rapach, Erik Christian Montes Schütte, and Sander Schwenk-Nebbe, which is available to download for free at SSRN: https://ssrn.com/abstract=4278745.
The $\text{PBSV}_p$ is a Shapley-based decomposition that measures the contributions of an individual predictor $p$ in fitted models to the out-of-sample loss. While a performance metric like the RMSE focuses solely on out-of-sample performance, MAS evaluates whether a model's out-of-sample success mirrors what it has learned from the in-sample data by comparing $\text{iShapley-VI}_p$ (or $\text{oShapley-VI}_p$) to $\text{PBSV}_p$. The MAS paired with a performance metric such as the RMSE provides insight into the model's "intentional success" (see below example).
The interpretation of PBSVs is straightforward: if $\text{PBSV}_p$ is negative (positive), predictor $p$ reduces (increases) the loss and is thus beneficial for (detrimental to) forecasting accuracy in the out-of-sample period. Taking the sum of the individual contributions (including the contribution of the empty set) yields the decomposed loss exactly (due to the efficiency property of Shapley values; see below example).
Please cite our paper if you find the package useful:
Borup, Daniel and Coulombe, Philippe Goulet and Rapach, David E. and Montes Schütte, Erik Christian and Schwenk-Nebbe, Sander (2022). “The Anatomy of Out-of-Sample Forecasting Accuracy”. Federal Reserve Bank of Atlanta Working Paper 2022-16. https://doi.org/10.29338/wp2022-16.
Quickstart
If you haven't already, install the package via pip install anatomy
, preferably in a new environment with Python 3.9.
The anatomy package uses a simple workflow. An Anatomy
object is initially estimated on your forecasting setup (using your data and your models), is then stored to disk, and can then be loaded at any future time without requiring re-estimation.
After initial estimation, an Anatomy
can anatomize:
- forecasts produced by any combination of your models
- your original forecasts
- any loss or gain function applied to your forecasts
- an arbitrary subset of your forecasts
all of which requires no additional computational time.
General structure
You may already have trained your models before you create the Anatomy
, and the aggregate of all your models at all periods may be too large to fit into your RAM. During estimation, the Anatomy
will therefore ask you for the specific model and dataset it needs at a given iteration by calling your mapping function:
from anatomy import *
def my_map(key: AnatomyModelProvider.PeriodKey) -> \
AnatomyModelProvider.PeriodValue:
train, test, model = ... # load from somewhere or generate here
return AnatomyModelProvider.PeriodValue(train, test, model)
You wrap the mapping function in an AnatomyModelProvider
alongside information about the forecasting application:
my_provider = AnatomyModelProvider(
n_periods=..., n_features=..., model_names=[...],
y_name=..., provider_fn=my_map
)
and finally create the Anatomy
:
my_anatomy = Anatomy(provider=my_provider, n_iterations=...).precompute(
n_jobs=16, save_path="my_anatomy.bin"
)
After running the above, the Anatomy
is estimated and stored in your working directory as my_anatomy.bin
.
Example:
For convenience, the examples below are contained in a single Python script available here.
To get started, we need a forecasting application. We use a linear DGP to generate our dataset consisting of 500 observations of the three predictors x_{0,1,2}
and our target y
:
# set random seed for reproducibility:
np.random.seed(1338)
xy = pd.DataFrame(np.random.normal(0, 1, (500, 3)), columns=["x_0", "x_1", "x_2"])
xy["y"] = xy.sum(axis=1) + np.random.normal(0, 1, 500)
# set a unique and monotonically increasing index (default index would suffice):
xy.index = pd.date_range("2021-04-19", "2022-08-31").map(lambda x: x.date())
For convenience, the AnatomySubsets
includes a generator that splits your dataset into training and test sets according to your forecasting scheme. Here, we include 100 periods in our first training set, forecast the target of the next period with no gap between the training set and the forecast, extend our training set by one period, and repeat until we reach the end of our data:
subsets = AnatomySubsets.generate(
index=xy.index,
initial_window=100,
estimation_type=AnatomySubsets.EstimationType.EXPANDING,
periods=1,
gap=0
)
In this example, we have not yet trained our models. We thus do so directly in our mapping function:
def mapper(key: AnatomyModelProvider.PeriodKey) -> \
AnatomyModelProvider.PeriodValue:
train = xy.iloc[subsets.get_train_subset(key.period)]
test = xy.iloc[subsets.get_test_subset(key.period)]
if key.model_name == "ols":
model = train_ols(train.drop("y", axis=1), train["y"])
elif key.model_name == "rf":
model = train_rf(train.drop("y", axis=1), train["y"])
return AnatomyModelProvider.PeriodValue(train, test, model)
using train_ols
and train_rf
, which train a model and yield its prediction function wrapped in an AnatomyModel
:
from sklearn.ensemble import RandomForestRegressor
from sklearn.linear_model import LinearRegression
def train_ols(x_train: pd.DataFrame, y_train: pd.Series) -> AnatomyModel:
ols_model = LinearRegression().fit(x_train, y_train)
def pred_fn_ols(xs: np.ndarray) -> np.ndarray:
xs_df = pd.DataFrame(xs, columns=x_train.columns)
return np.array(ols_model.predict(xs_df)).flatten()
return AnatomyModel(pred_fn_ols)
def train_rf(x_train: pd.DataFrame, y_train: pd.Series) -> AnatomyModel:
rf_model = RandomForestRegressor(random_state=1338).fit(x_train, y_train)
def pred_fn_rf(xs: np.ndarray) -> np.ndarray:
xs_df = pd.DataFrame(xs, columns=x_train.columns)
return np.array(rf_model.predict(xs_df)).flatten()
return AnatomyModel(pred_fn_rf)
We now have all we need to train the models and estimate the Anatomy
:
provider = AnatomyModelProvider(
n_periods=subsets.n_periods,
n_features=xy.shape[1]-1,
model_names=["ols", "rf"],
y_name="y",
provider_fn=mapper
)
anatomy = Anatomy(provider=provider, n_iterations=10).precompute(
n_jobs=16, save_path="anatomy.bin"
)
At this point, the Anatomy
is stored as anatomy.bin
in our working directory. We can load it at any later point using anatomy = Anatomy.load("anatomy.bin")
.
Anatomizing
We can now use our estimated Anatomy
to anatomize our forecasts. In this example, we are using two models, rf
and ols
, as well as an equal-weighted combination of the two:
groups = {
"rf": ["rf"],
"ols": ["ols"],
"ols+rf": ["ols", "rf"]
}
Anatomize the out-of-sample $R^2$ of the forecasts:
To decompose the out-of-sample $R^2$ of our forecasts produced by the two models and their combination, we use the unconditional forecasts as benchmark and provide a function transforming forecasts into out-of-sample $R^2$ to the Anatomy
:
prevailing_mean = np.array([
xy.iloc[subsets.get_train_subset(period=i)]["y"].mean()
for i in range(subsets.n_periods)
])
def transform(y_hat, y):
return 1 - np.sum((y - y_hat) ** 2) / np.sum((y - prevailing_mean) ** 2)
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
This yields the change in out-of-sample $R^2$ attributable to each predictor:
>>> df
base_contribution x_0 x_1 x_2
rf 2021-07-28 -> 2022-08-31 0.000759 0.257240 0.238458 0.215062
ols 2021-07-28 -> 2022-08-31 0.000000 0.285329 0.267940 0.249870
ols+rf 2021-07-28 -> 2022-08-31 0.000383 0.279073 0.259201 0.240548
Interpretation
The Shapley-based decomposition can be understood as a means to fairly allocate a single value (in this case the out-of-sample $R^2$) amongst multiple actors contributiong to it (the predictors in our model). This implies that the individual contributions of the actors and the contribution of the empty set of actors (base_contribution
) sum up exactly to the original value that was decomposed.
The above depicts the individual contributions to the out-of-sample $R^2$, which can be negative, if a given predictor hurts accuracy, or positive, if a given predictor increases accuracy. In this case, all predictors contribute positively to the out-of-sample $R^2$. In practice, predictors can hurt accuracy by reducing the out-of-sample $R^2$.
Note: In this example, we use the prevailing mean (average of the target in the training sets) as benchmark to compute the out-of-sample R². The average forecast of an OLS model coincides with this benchmark, which explains why the base_contribution
of OLS is exactly zero.
... the RMSE of the forecasts:
def transform(y_hat, y):
return np.sqrt(np.mean((y - y_hat) ** 2))
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
which yields the change in root mean squared error attributable to each predictor:
>>> df
base_contribution x_0 x_1 x_2
rf 2021-07-28 -> 2022-08-31 2.105513 -0.351430 -0.326065 -0.296708
ols 2021-07-28 -> 2022-08-31 2.106313 -0.414488 -0.386125 -0.371149
ols+rf 2021-07-28 -> 2022-08-31 2.105910 -0.397903 -0.368193 -0.350080
Interpretation
We previously decomposed the out-of-sample $R^2$. In this case, we use the RMSE loss function, implying that a predictor with a negative contribution increases forecasting accuracy. Because the RMSE cannot be negative, the base_contribution
, which is the RMSE of the average forecasts of the models, can only be positive.
Similar to the previous case, we find that all predictors contribute positively to forecasting accuracy (by contributing negatively to the RMSE).
... the MAE:
def transform(y_hat, y):
return np.mean(np.abs(y - y_hat))
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
which yields the change in mean absolute error attributable to each predictor:
>>> df
base_contribution x_0 x_1 x_2
rf 2021-07-28 -> 2022-08-31 1.679359 -0.299382 -0.249591 -0.221651
ols 2021-07-28 -> 2022-08-31 1.679946 -0.345583 -0.300613 -0.288960
ols+rf 2021-07-28 -> 2022-08-31 1.679652 -0.330303 -0.283129 -0.262270
Interpretation
The interpretation is similar to that of the RMSE.
... the SE:
def transform(y_hat, y):
return (y - y_hat) ** 2
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
which yields the change in squared error attributable to each predictor for each forecast:
>>> df
base_contribution x_0 x_1 x_2
rf 2021-07-28 0.026485 -0.063311 -0.115985 0.163743
2021-07-29 0.021582 0.238569 -0.370448 0.694773
2021-07-30 2.451742 -2.702365 1.660915 -1.407192
... ... ... ... ...
Note: The transform
function in this case does not aggregate (returns a vector instead of a scalar). The Anatomy
thus yields one decomposition per forecast, which is also known as a local (as opposed to global) decomposition.
Interpretation
The previous decompositions have consistently shown that all predictors increase forecasting accuracy when it is gauged over the entire period (2021-07-28 to 2022-08-31). Anatomizing instead each individual forecast reveals that this is not always true, at least not at the local level. We now see that individual predictors are contributing positively to the squared error of some forecasts (thus reducing forecast accuracy).
... the RMSE of the forecasts in a subperiod:
subset = pd.date_range("2021-07-28", "2021-08-06").map(lambda x: x.date())
def transform(y_hat, y):
return np.sqrt(np.mean((y - y_hat) ** 2))
df_pbsv_rmse = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform),
explanation_subset=subset
)
which yields the change in root mean squared error in the ten-day period attributable to each predictor:
>>> df_pbsv_rmse
base_contribution x_0 x_1 x_2
rf 2021-07-28 -> 2021-08-06 1.402950 -0.149847 -0.199615 0.118059
ols 2021-07-28 -> 2021-08-06 1.404761 -0.155275 -0.323760 0.317055
ols+rf 2021-07-28 -> 2021-08-06 1.403853 -0.169584 -0.275408 0.211672
Interpretation
In this short subperiod of ten days, we find that our last predictor contributed positively to the RMSE (and thus negatively to forecasting accuracy).
... or just the raw forecasts:
def transform(y_hat):
return y_hat
df_oshapley = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
which yields the change in the forecast attributable to each predictor:
>>> df_oshapley
base_contribution x_0 x_1 x_2
rf 2021-07-28 0.070861 0.222932 0.360182 -0.315820
2021-07-29 0.070354 0.250389 -0.617779 0.984992
2021-07-30 0.071163 -1.514547 0.839562 -0.835136
... ... ... ... ...
Interpretation
Decomposing the forecasts themselves yields contributions that bear no relation to forecasting accuracy. Hence, a negative or positive contribution means no more than a decrease or increase in the forecast at that period attributable to the given predictor, which may or may not have be good for forecasting accuracy.
Hence: beware, a high average absolute contribution does not necessarily translate into a high gain in accuracy. That is precisely why we need to decompose the loss directly, and in consequence, take into account our target and how far away our forecasts were from it.
From the anatomized raw forecasts, we can compute the $\text{oShapley-VI}$ by averaging over the magnitudes of the invididual contributions (here for the ols+rf
combination):
>>> df_oshapley.loc["ols+rf"].abs().mean(axis=0)
base_contribution 0.078415
x_0 0.864057
x_1 0.786044
x_2 0.770115
The Efficiency property
During estimation, the Anatomy
checks that the individual attributions of the predictors to the forecasts sum up exactly to the forecasts produced by the models. The estimation would be aborted if efficiency does not hold.
Due to the efficiency property of Shapley values, summing the individual contributions yields the decomposed value exactly. We can check that the results that Anatomy
yields are consistet. We can recover the RMSE from the decomposed RMSE, but we can also recover the RMSE from the decomposed forecasts. Let's make sure that they match.
We first anatomize the RMSE:
def transform(y_hat, y):
return np.sqrt(np.mean((y - y_hat) ** 2))
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
The RMSE (or any other decomposed value) is the sum of the individual attributions, rmse_a = df.sum(axis=1)
:
>>> rmse_a
rf 2021-07-28 -> 2022-08-31 1.131310
ols 2021-07-28 -> 2022-08-31 0.934551
ols+rf 2021-07-28 -> 2022-08-31 0.989733
We next decompose the raw forecasts and compute the RMSE from these:
def transform(y_hat):
return y_hat
df = anatomy.explain(
model_sets=AnatomyModelCombination(groups=groups),
transformer=AnatomyModelOutputTransformer(transform=transform)
)
The forecasts are recovered as the sum of the individual attributions, y_hat = df.sum(axis=1)
:
>>> y_hat
rf 2021-07-28 0.338154
2021-07-29 0.687956
2021-07-30 -1.438958
...
From the forecasts, we can compute the RMSE:
y_true = np.hstack([
xy.iloc[subsets.get_test_subset(period=i)]["y"]
for i in range(subsets.n_periods)
])
rmse_b = pd.Series({
key: np.sqrt(np.mean((y_true - y_hat.xs(key)) ** 2))
for key in groups.keys()
})
which yields the same RMSE as the sum of the contributions of the RMSE decomposition:
>>> rmse_b
rf 1.131310
ols 0.934551
ols+rf 0.989733
Model Accordance Score:
We can compute the MAS for our combination model (ols+rf
) using $\text{oShapley-VI}$ and, for instance, the
$\text{PBSV}$ for the root-mean-square error:
vi = df_oshapley.loc["ols+rf"].abs().mean(axis=0)
pbsv = df_pbsv_rmse.loc["ols+rf"].iloc[0]
# loss type is rmse, so lower is better:
loss_type = MAS.LossType.LOWER_IS_BETTER
mas = MAS(vi, pbsv, loss_type).compute()
which yields the MAS:
>>> mas
{'mas': 1.0, 'mas_p_value': 0.020796}
Interpretation
In this example, the ranking of $\text{oShapley-VI}$ is identical to the signed-ranking of $\text{PBSV}$.
Thus, MAS is 1 (perfect) and the null hypothesis of no relation between $\text{oShapley-VI}_p$ and $\text{PBSV}_p$
is rejected at the 5% level (mas_p_value
is the probability of observing a MAS at least as extreme as mas
under
the null).
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.