| 1 | /////////////////////////////////////////////////////////////////////////////// |
| 2 | // tail_mean.hpp |
| 3 | // |
| 4 | // Copyright 2006 Daniel Egloff, Olivier Gygi. Distributed under the Boost |
| 5 | // Software License, Version 1.0. (See accompanying file |
| 6 | // LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt) |
| 7 | |
| 8 | #ifndef BOOST_ACCUMULATORS_STATISTICS_TAIL_MEAN_HPP_DE_01_01_2006 |
| 9 | #define BOOST_ACCUMULATORS_STATISTICS_TAIL_MEAN_HPP_DE_01_01_2006 |
| 10 | |
| 11 | #include <numeric> |
| 12 | #include <vector> |
| 13 | #include <limits> |
| 14 | #include <functional> |
| 15 | #include <sstream> |
| 16 | #include <stdexcept> |
| 17 | #include <boost/throw_exception.hpp> |
| 18 | #include <boost/parameter/keyword.hpp> |
| 19 | #include <boost/mpl/placeholders.hpp> |
| 20 | #include <boost/type_traits/is_same.hpp> |
| 21 | #include <boost/accumulators/framework/accumulator_base.hpp> |
| 22 | #include <boost/accumulators/framework/extractor.hpp> |
| 23 | #include <boost/accumulators/numeric/functional.hpp> |
| 24 | #include <boost/accumulators/framework/parameters/sample.hpp> |
| 25 | #include <boost/accumulators/statistics_fwd.hpp> |
| 26 | #include <boost/accumulators/statistics/count.hpp> |
| 27 | #include <boost/accumulators/statistics/tail.hpp> |
| 28 | #include <boost/accumulators/statistics/tail_quantile.hpp> |
| 29 | #include <boost/accumulators/statistics/parameters/quantile_probability.hpp> |
| 30 | |
| 31 | #ifdef _MSC_VER |
| 32 | # pragma warning(push) |
| 33 | # pragma warning(disable: 4127) // conditional expression is constant |
| 34 | #endif |
| 35 | |
| 36 | namespace boost { namespace accumulators |
| 37 | { |
| 38 | |
| 39 | namespace impl |
| 40 | { |
| 41 | |
| 42 | /////////////////////////////////////////////////////////////////////////////// |
| 43 | // coherent_tail_mean_impl |
| 44 | // |
| 45 | /** |
| 46 | @brief Estimation of the coherent tail mean based on order statistics (for both left and right tails) |
| 47 | |
| 48 | The coherent tail mean \f$\widehat{CTM}_{n,\alpha}(X)\f$ is equal to the non-coherent tail mean \f$\widehat{NCTM}_{n,\alpha}(X)\f$ |
| 49 | plus a correction term that ensures coherence in case of non-continuous distributions. |
| 50 | |
| 51 | \f[ |
| 52 | \widehat{CTM}_{n,\alpha}^{\mathrm{right}}(X) = \widehat{NCTM}_{n,\alpha}^{\mathrm{right}}(X) + |
| 53 | \frac{1}{\lceil n(1-\alpha)\rceil}\hat{q}_{n,\alpha}(X)\left(1 - \alpha - \frac{1}{n}\lceil n(1-\alpha)\rceil \right) |
| 54 | \f] |
| 55 | |
| 56 | \f[ |
| 57 | \widehat{CTM}_{n,\alpha}^{\mathrm{left}}(X) = \widehat{NCTM}_{n,\alpha}^{\mathrm{left}}(X) + |
| 58 | \frac{1}{\lceil n\alpha\rceil}\hat{q}_{n,\alpha}(X)\left(\alpha - \frac{1}{n}\lceil n\alpha\rceil \right) |
| 59 | \f] |
| 60 | */ |
| 61 | template<typename Sample, typename LeftRight> |
| 62 | struct coherent_tail_mean_impl |
| 63 | : accumulator_base |
| 64 | { |
| 65 | typedef typename numeric::functional::fdiv<Sample, std::size_t>::result_type float_type; |
| 66 | // for boost::result_of |
| 67 | typedef float_type result_type; |
| 68 | |
| 69 | coherent_tail_mean_impl(dont_care) {} |
| 70 | |
| 71 | template<typename Args> |
| 72 | result_type result(Args const &args) const |
| 73 | { |
| 74 | std::size_t cnt = count(args); |
| 75 | |
| 76 | std::size_t n = static_cast<std::size_t>( |
| 77 | std::ceil( |
| 78 | cnt * ( ( is_same<LeftRight, left>::value ) ? args[quantile_probability] : 1. - args[quantile_probability] ) |
| 79 | ) |
| 80 | ); |
| 81 | |
| 82 | extractor<tag::non_coherent_tail_mean<LeftRight> > const some_non_coherent_tail_mean = {}; |
| 83 | |
| 84 | return some_non_coherent_tail_mean(args) |
| 85 | + numeric::fdiv(quantile(args), n) |
| 86 | * ( |
| 87 | ( is_same<LeftRight, left>::value ) ? args[quantile_probability] : 1. - args[quantile_probability] |
| 88 | - numeric::fdiv(n, count(args)) |
| 89 | ); |
| 90 | } |
| 91 | |
| 92 | // serialization is done by accumulators it depends on |
| 93 | template<class Archive> |
| 94 | void serialize(Archive & ar, const unsigned int file_version) {} |
| 95 | }; |
| 96 | |
| 97 | /////////////////////////////////////////////////////////////////////////////// |
| 98 | // non_coherent_tail_mean_impl |
| 99 | // |
| 100 | /** |
| 101 | @brief Estimation of the (non-coherent) tail mean based on order statistics (for both left and right tails) |
| 102 | |
| 103 | An estimation of the non-coherent tail mean \f$\widehat{NCTM}_{n,\alpha}(X)\f$ is given by the mean of the |
| 104 | \f$\lceil n\alpha\rceil\f$ smallest samples (left tail) or the mean of the \f$\lceil n(1-\alpha)\rceil\f$ |
| 105 | largest samples (right tail), \f$n\f$ being the total number of samples and \f$\alpha\f$ the quantile level: |
| 106 | |
| 107 | \f[ |
| 108 | \widehat{NCTM}_{n,\alpha}^{\mathrm{right}}(X) = \frac{1}{\lceil n(1-\alpha)\rceil} \sum_{i=\lceil \alpha n \rceil}^n X_{i:n} |
| 109 | \f] |
| 110 | |
| 111 | \f[ |
| 112 | \widehat{NCTM}_{n,\alpha}^{\mathrm{left}}(X) = \frac{1}{\lceil n\alpha\rceil} \sum_{i=1}^{\lceil \alpha n \rceil} X_{i:n} |
| 113 | \f] |
| 114 | |
| 115 | It thus requires the caching of at least the \f$\lceil n\alpha\rceil\f$ smallest or the \f$\lceil n(1-\alpha)\rceil\f$ |
| 116 | largest samples. |
| 117 | |
| 118 | @param quantile_probability |
| 119 | */ |
| 120 | template<typename Sample, typename LeftRight> |
| 121 | struct non_coherent_tail_mean_impl |
| 122 | : accumulator_base |
| 123 | { |
| 124 | typedef typename numeric::functional::fdiv<Sample, std::size_t>::result_type float_type; |
| 125 | // for boost::result_of |
| 126 | typedef float_type result_type; |
| 127 | |
| 128 | non_coherent_tail_mean_impl(dont_care) {} |
| 129 | |
| 130 | template<typename Args> |
| 131 | result_type result(Args const &args) const |
| 132 | { |
| 133 | std::size_t cnt = count(args); |
| 134 | |
| 135 | std::size_t n = static_cast<std::size_t>( |
| 136 | std::ceil( |
| 137 | cnt * ( ( is_same<LeftRight, left>::value ) ? args[quantile_probability] : 1. - args[quantile_probability] ) |
| 138 | ) |
| 139 | ); |
| 140 | |
| 141 | // If n is in a valid range, return result, otherwise return NaN or throw exception |
| 142 | if (n <= static_cast<std::size_t>(tail(args).size())) |
| 143 | return numeric::fdiv( |
| 144 | std::accumulate( |
| 145 | tail(args).begin() |
| 146 | , tail(args).begin() + n |
| 147 | , Sample(0) |
| 148 | ) |
| 149 | , n |
| 150 | ); |
| 151 | else |
| 152 | { |
| 153 | if (std::numeric_limits<result_type>::has_quiet_NaN) |
| 154 | { |
| 155 | return std::numeric_limits<result_type>::quiet_NaN(); |
| 156 | } |
| 157 | else |
| 158 | { |
| 159 | std::ostringstream msg; |
| 160 | msg << "index n = " << n << " is not in valid range [0, " << tail(args).size() << ")" ; |
| 161 | boost::throw_exception(e: std::runtime_error(msg.str())); |
| 162 | return Sample(0); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | // serialization is done by accumulators it depends on |
| 168 | template<class Archive> |
| 169 | void serialize(Archive & ar, const unsigned int file_version) {} |
| 170 | }; |
| 171 | |
| 172 | } // namespace impl |
| 173 | |
| 174 | |
| 175 | /////////////////////////////////////////////////////////////////////////////// |
| 176 | // tag::coherent_tail_mean<> |
| 177 | // tag::non_coherent_tail_mean<> |
| 178 | // |
| 179 | namespace tag |
| 180 | { |
| 181 | template<typename LeftRight> |
| 182 | struct coherent_tail_mean |
| 183 | : depends_on<count, quantile, non_coherent_tail_mean<LeftRight> > |
| 184 | { |
| 185 | typedef accumulators::impl::coherent_tail_mean_impl<mpl::_1, LeftRight> impl; |
| 186 | }; |
| 187 | |
| 188 | template<typename LeftRight> |
| 189 | struct non_coherent_tail_mean |
| 190 | : depends_on<count, tail<LeftRight> > |
| 191 | { |
| 192 | typedef accumulators::impl::non_coherent_tail_mean_impl<mpl::_1, LeftRight> impl; |
| 193 | }; |
| 194 | |
| 195 | struct abstract_non_coherent_tail_mean |
| 196 | : depends_on<> |
| 197 | { |
| 198 | }; |
| 199 | } |
| 200 | |
| 201 | /////////////////////////////////////////////////////////////////////////////// |
| 202 | // extract::non_coherent_tail_mean; |
| 203 | // extract::coherent_tail_mean; |
| 204 | // |
| 205 | namespace extract |
| 206 | { |
| 207 | extractor<tag::abstract_non_coherent_tail_mean> const = {}; |
| 208 | extractor<tag::tail_mean> const = {}; |
| 209 | |
| 210 | BOOST_ACCUMULATORS_IGNORE_GLOBAL(non_coherent_tail_mean) |
| 211 | BOOST_ACCUMULATORS_IGNORE_GLOBAL(coherent_tail_mean) |
| 212 | } |
| 213 | |
| 214 | using extract::non_coherent_tail_mean; |
| 215 | using extract::coherent_tail_mean; |
| 216 | |
| 217 | // for the purposes of feature-based dependency resolution, |
| 218 | // coherent_tail_mean<LeftRight> provides the same feature as tail_mean |
| 219 | template<typename LeftRight> |
| 220 | struct feature_of<tag::coherent_tail_mean<LeftRight> > |
| 221 | : feature_of<tag::tail_mean> |
| 222 | { |
| 223 | }; |
| 224 | |
| 225 | template<typename LeftRight> |
| 226 | struct feature_of<tag::non_coherent_tail_mean<LeftRight> > |
| 227 | : feature_of<tag::abstract_non_coherent_tail_mean> |
| 228 | { |
| 229 | }; |
| 230 | |
| 231 | // So that non_coherent_tail_mean can be automatically substituted |
| 232 | // with weighted_non_coherent_tail_mean when the weight parameter is non-void. |
| 233 | template<typename LeftRight> |
| 234 | struct as_weighted_feature<tag::non_coherent_tail_mean<LeftRight> > |
| 235 | { |
| 236 | typedef tag::non_coherent_weighted_tail_mean<LeftRight> type; |
| 237 | }; |
| 238 | |
| 239 | template<typename LeftRight> |
| 240 | struct feature_of<tag::non_coherent_weighted_tail_mean<LeftRight> > |
| 241 | : feature_of<tag::non_coherent_tail_mean<LeftRight> > |
| 242 | {}; |
| 243 | |
| 244 | // NOTE that non_coherent_tail_mean cannot be feature-grouped with tail_mean, |
| 245 | // which is the base feature for coherent tail means, since (at least for |
| 246 | // non-continuous distributions) non_coherent_tail_mean is a different measure! |
| 247 | |
| 248 | }} // namespace boost::accumulators |
| 249 | |
| 250 | #ifdef _MSC_VER |
| 251 | # pragma warning(pop) |
| 252 | #endif |
| 253 | |
| 254 | #endif |
| 255 | |