def test_segment_diffs_no_segment_exclusive(self): x = tf.constant([11, 13, 17, 19, 23]) dx1 = self.evaluate( segment_ops.segment_diff(x, segment_ids=None, order=1, exclusive=True)) np.testing.assert_array_equal(dx1, [2, 4, 2, 4]) dx2 = self.evaluate( segment_ops.segment_diff(x, segment_ids=None, order=2, exclusive=True)) np.testing.assert_array_equal(dx2, [6, 6, 6])
def test_segment_diffs_segment_exclusive(self): x = tf.constant([2, 5, 1, 7, 9] + [32, 10, 12, 3] + [4, 8, 5]) segments = tf.constant([0, 0, 0, 0, 0] + [1, 1, 1, 1] + [2, 2, 2]) dx1 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=1, exclusive=True)) np.testing.assert_array_equal(dx1, ([3, -4, 6, 2] + [-22, 2, -9] + [4, -3])) dx2 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=2, exclusive=True)) np.testing.assert_array_equal(dx2, ([-1, 2, 8] + [-20, -7] + [1]))
def test_segment_diffs_small_segment(self): x = tf.constant([2, 5, 1, 7] + [9] + [32, 10, 12, 3] + [4, 8, 5]) segments = tf.constant([0, 0, 0, 0] + [1] + [2, 2, 2, 2] + [3, 3, 3]) dx1 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=2, exclusive=False)) np.testing.assert_array_equal( dx1, ([2, 5, -1, 2] + [9] + [32, 10, -20, -7] + [4, 8, 1])) dx2 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=2, exclusive=True)) np.testing.assert_array_equal(dx2, ([-1, 2] + [] + [-20, -7] + [1]))
def test_segment_diffs_large_order(self): x = tf.constant([2, 5, 1, 7, 9] + [32, 10, 12, 3] + [4, 8, 5]) segments = tf.constant([0, 0, 0, 0, 0] + [1, 1, 1, 1] + [2, 2, 2]) dx1 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=4, exclusive=False)) np.testing.assert_array_equal( dx1, ([2, 5, 1, 7, 7] + [32, 10, 12, 3] + [4, 8, 5])) dx2 = self.evaluate( segment_ops.segment_diff( x, segment_ids=segments, order=4, exclusive=True)) np.testing.assert_array_equal( dx2, ([7] + [] + [])) # The empty arrays are for the segments 1 and 2.
def forward_rates_from_yields(yields, times, groups=None, dtype=None, name=None): """Computes forward rates given a set of zero rates. Denote the price of a zero coupon bond maturing at time `t` by `Z(t)`. Then the zero rate to time `t` is defined as ```None r(t) = - ln(Z(t)) / t (1) ``` This is the (continuously compounded) interest rate that applies between time `0` and time `t` as seen at time `0`. The forward rate between times `t1` and `t2` is defined as the interest rate that applies to the period `[t1, t2]` as seen from today. It is related to the zero coupon bond prices by ```None exp(-f(t1, t2)(t2-t1)) = Z(t2) / Z(t1) (2) f(t1, t2) = - (ln Z(t2) - ln Z(t1)) / (t2 - t1) (3) f(t1, t2) = (t2 * r(t2) - t1 * r(t1)) / (t2 - t1) (4) ``` Given a sequence of increasing times `[t1, t2, ... tn]` and the zero rates for those times, this function computes the forward rates that apply to the consecutive time intervals i.e. `[0, t1], [t1, t2], ... [t_{n-1}, tn]` using Eq. (4) above. Note that for the interval `[0, t1]` the forward rate is the same as the zero rate. Additionally, this function supports this computation for a batch of such rates. Batching is made slightly complicated by the fact that different zero curves may have different numbers of tenors (the parameter `n` above). Instead of a batch as an extra dimension, we support the concept of groups (also see documentation for `tf.segment_sum` which uses the same concept). ### Example The following example illustrates this method along with the concept of groups. Assuming there are two sets of zero rates (e.g. for different currencies) whose implied forward rates are needed. The first set has a total of three marked tenors at `[0.25, 0.5, 1.0]`. The second set has four marked tenors at `[0.25, 0.5, 1.0, 1.5]`. Suppose, the zero rates for the first set are: `[0.04, 0.041, 0.044]` and the second are `[0.022, 0.025, 0.028, 0.036]`. Then this data is batched together as follows: Groups: [0, 0 0, 1, 1, 1 1 ] First three times for group 0, next four for group 1. Times: [0.25, 0.5, 1.0, 0.25, 0.5, 1.0, 1.5] First three rates for group 0, next four for group 1. Rates: [0.04, 0.041, 0.044, 0.022, 0.025, 0.028, 0.036] ```python dtype = np.float64 groups = np.array([0, 0 0, 1, 1, 1, 1 ]) times = np.array([0.25, 0.5, 1.0, 0.25, 0.5, 1.0, 1.5], dtype=dtype) rates = np.array([0.04, 0.041, 0.044, 0.022, 0.025, 0.028, 0.036], dtype=dtype) forward_rates = forward_rates_from_yields( rates, times, groups=groups, dtype=dtype) ``` ### References: [1]: John C. Hull. Options, Futures and Other Derivatives. Ninth Edition. June 2006. Args: yields: Real rank 1 `Tensor` of size `n`. The discount/zero rates. times: Real positive rank 1 `Tensor` of size `n`. The set of times corresponding to the supplied zero rates. If no `groups` is supplied, then the whole array should be sorted in an increasing order. If `groups` are supplied, then the times within a group should be in an increasing order. groups: Optional int `Tensor` of size `n` containing values between 0 and `k-1` where `k` is the number of different curves. Default value: None. This implies that all the rates are treated as a single group. dtype: `tf.Dtype`. If supplied the dtype for the `yields` and `times`. Default value: None which maps to the default dtype inferred by Tensorflow (float32). name: Python str. The name to give to the ops created by this function. Default value: None which maps to 'forward_rates_from_yields'. Returns: Real rank 1 `Tensor` of size `n` containing the forward rate that applies for each successive time interval (within each group if groups are specified). """ with tf.compat.v1.name_scope(name, default_name='forward_rates_from_yields', values=[yields, times, groups]): yields = tf.convert_to_tensor(yields, dtype=dtype, name='yields') times = tf.convert_to_tensor(times, dtype=dtype, name='times') if groups is not None: groups = tf.convert_to_tensor(groups, name='groups') # (t2 * r(t2) - t1 * r(t1)) / (t2 - t1) rate_times = yields * times diff_rate_times = segment_ops.segment_diff(rate_times, order=1, exclusive=False, segment_ids=groups) diff_times = segment_ops.segment_diff(times, order=1, exclusive=False, segment_ids=groups) return diff_rate_times / diff_times
def yields_from_forward_rates(discrete_forwards, times, groups=None, dtype=None, name=None): """Computes yield rates from discrete forward rates. Denote the price of a zero coupon bond maturing at time `t` by `Z(t)`. Then the zero rate to time `t` is defined as ```None r(t) = - ln(Z(t)) / t (1) ``` This is the (continuously compounded) interest rate that applies between time `0` and time `t` as seen at time `0`. The forward rate between times `t1` and `t2` is defined as the interest rate that applies to the period `[t1, t2]` as seen from today. It is related to the zero coupon bond prices by ```None exp(-f(t1, t2)(t2-t1)) = Z(t2) / Z(t1) (2) f(t1, t2) = - (ln Z(t2) - ln Z(t1)) / (t2 - t1) (3) f(t1, t2) = (t2 * r(t2) - t1 * r(t1)) / (t2 - t1) (4) ``` Given a sequence of increasing times `[t1, t2, ... tn]` and the forward rates for the consecutive time intervals, i.e. `[0, t1]`, `[t1, t2]` to `[t_{n-1}, tn]`, this function computes the yields to maturity for maturities `[t1, t2, ... tn]` using Eq. (4) above. Additionally, this function supports this computation for a batch of such forward rates. Batching is made slightly complicated by the fact that different zero curves may have different numbers of tenors (the parameter `n` above). Instead of a batch as an extra dimension, we support the concept of groups (also see documentation for `tf.segment_sum` which uses the same concept). ### Example The following example illustrates this method along with the concept of groups. Assuming there are two sets of zero rates (e.g. for different currencies) whose implied forward rates are needed. The first set has a total of three marked tenors at `[0.25, 0.5, 1.0]`. The second set has four marked tenors at `[0.25, 0.5, 1.0, 1.5]`. Suppose, the forward rates for the first set are: `[0.04, 0.041, 0.044]` and the second are `[0.022, 0.025, 0.028, 0.036]`. Then this data is batched together as follows: Groups: [0, 0 0, 1, 1, 1 1 ] First three times for group 0, next four for group 1. Times: [0.25, 0.5, 1.0, 0.25, 0.5, 1.0, 1.5] First three discrete forwards for group 0, next four for group 1. Forwards: [0.04, 0.042, 0.047, 0.022, 0.028, 0.031, 0.052] ```python dtype = np.float64 groups = np.array([0, 0 0, 1, 1, 1, 1 ]) times = np.array([0.25, 0.5, 1.0, 0.25, 0.5, 1.0, 1.5], dtype=dtype) discrete_forwards = np.array( [0.04, 0.042, 0.047, 0.022, 0.028, 0.031, 0.052], dtype=dtype) yields = yields_from_forward_rates(discrete_forwards, times, groups=groups, dtype=dtype) # Produces: [0.04, 0.041, 0.044, 0.022, 0.025, 0.028, 0.036] ``` ### References: [1]: John C. Hull. Options, Futures and Other Derivatives. Ninth Edition. June 2006. Args: discrete_forwards: Real rank 1 `Tensor` of size `n`. The forward rates for the time periods specified. Note that the first value applies between `0` and time `times[0]`. times: Real positive rank 1 `Tensor` of size `n`. The set of times corresponding to the supplied zero rates. If no `groups` is supplied, then the whole array should be sorted in an increasing order. If `groups` are supplied, then the times within a group should be in an increasing order. groups: Optional int `Tensor` of size `n` containing values between 0 and `k-1` where `k` is the number of different curves. Default value: None. This implies that all the rates are treated as a single group. dtype: `tf.Dtype`. If supplied the dtype for the `discrete_forwards` and `times`. Default value: None which maps to the default dtype inferred by Tensorflow. name: Python str. The name to give to the ops created by this function. Default value: None which maps to 'yields_from_forward_rates'. Returns: yields: Real rank 1 `Tensor` of size `n` containing the zero coupon yields that for the supplied maturities (within each group if groups are specified). """ with tf.compat.v1.name_scope(name, default_name='yields_from_forward_rates', values=[discrete_forwards, times, groups]): discrete_forwards = tf.convert_to_tensor(discrete_forwards, dtype=dtype, name='discrete_forwards') times = tf.convert_to_tensor(times, dtype=dtype, name='times') if groups is not None: groups = tf.convert_to_tensor(groups, name='groups') # Strategy for solving this equation without loops. # Define x_i = f_i (t_i - t_{i-1}) where f are the forward rates and # t_{-1}=0. Also define y_i = r_i t_i # Then the relationship between the forward rate and the yield can be # written as: x_i = y_i - y_{i-1} which we need to solve for y. # Hence, y_i = x_0 + x_1 + ... x_i. intervals = segment_ops.segment_diff(times, order=1, exclusive=False, segment_ids=groups) x = intervals * discrete_forwards y = segment_ops.segment_cumsum(x, exclusive=False, segment_ids=groups) return y / times