ENH: Proof of Concept: Move NPV to Numba #91
Conversation
This seems to be more straight forward to maintain.
| @nb.jit(forceobj=True) # Need ``forceobj`` to support decimal.Decimal | ||
| def _npv_decimal(rates, cashflows, result): | ||
| r"""Version of the ``npv`` function supporting ``decimal.Decimal`` types | ||
|
|
||
| Warnings | ||
| -------- | ||
| For internal use only, note that this function performs no error checking. | ||
| """ | ||
| for i in range(rates.shape[0]): | ||
| for j in range(cashflows.shape[0]): | ||
| acc = Decimal("0.0") | ||
| for t in range(cashflows.shape[1]): | ||
| acc += cashflows[j, t] / ((Decimal("1.0") + rates[i]) ** t) | ||
| result[i, j] = acc |
There was a problem hiding this comment.
This is essentially a copy of _npv with the constants replaced with decimal.Decimals. I've tried doing something along the lines of dtype = Decimal if arr.dtype = np.dtype("O") else arr.dtype but this wouldn't compile in nopython mode. I'm trying to think of a neater way to implement this but haven't been able to think of one yet
| r"""Native version of the ``npv`` function. | ||
|
|
||
| Warnings | ||
| -------- | ||
| For internal use only, note that this function performs no error checking. |
There was a problem hiding this comment.
Documentation still needs to be improved; this is the workhorse powering the npv function.
| @@ -825,6 +827,40 @@ def irr(values, *, guess=None, tol=1e-12, maxiter=100, raise_exceptions=False): | |||
| return np.nan | |||
|
|
|||
|
|
|||
| @nb.njit(parallel=True) | |||
There was a problem hiding this comment.
I'm not sure if it's worth parallelising this function.
|
|
||
| r = np.atleast_1d(rate) | ||
| v = np.atleast_2d(values) | ||
|
|
There was a problem hiding this comment.
TODO: Add checking of array shapes if they are compatible.
| actual_npvs = npf.npv(0.05, cashflows) | ||
| assert_allclose(actual_npvs, expected_npvs) | ||
|
|
||
| @pytest.mark.parametrize("dtype", [Decimal, float]) |
There was a problem hiding this comment.
TODO: check with other dtypes e.g. float32
| for i, r in enumerate(rates): | ||
| for j, cf in enumerate(cashflows): | ||
| expected[i, j] = npf.npv(r, cf).item() |
There was a problem hiding this comment.
Here we're using a for loop to calculate the equivalent to the "broadcasting" operation. This is slow, however I am more confident that it is correct.
Kai-Striega
changed the title
|
Closing as Numba hasn't released Python 3.12 support yet. I'll also be exploring the Cython approach as an alternative approach. |
This PR moves the Net Present Value function to Numba
It is still a WIP, I'm not sure if the design decisions I've made are correct. Feedback is welcome.
I've tried using
numba.guvectorizeto make gufuncs out of the NumPy-Financial functions. This didn't work well (because we're not using "proper" gufuncs. Instead I've written them as for loops that are spead up with numba'sjit/njitfunctions.This leads to some duplication as we need both a
jitandnjitdecorator to supportdecimal.Decimal(which is a PyObject).Note that Numba does not support Python3.12 yet, this is expected to be out soon (in the 0.59 release) see the tracking issue for more details