Daily SOFR Swap Valuation: Separate Overnight Roll from Market P&L

# Calculating swap npv 1 day in the futures with rateslib


# Calculating swap npv 1 day in the futures with rateslib












I’m trying to evaluate a SOFR IRS at two different dates using RatesLib.

At date t, I build a curve using market data and use it to price an IRS.

At date t+1, I revalue the same swap using the updated market curve built from t+1 data — essentially computing the mark-to-market of the swap initiated at t.

My expectation is that, if this is done daily, the mark-to-market changes should approximately match the cumulative difference in swap rates multiplied by the duration, plus the carry. However, the results I obtain are quite noisy and significantly deviate from expectations.

Has anyone encountered a similar issue or found a possible explanation for this?

Here is what I am doing for a 10Y USD SOFR swap (curves is a dict containing my fitted curves):

```
    as_of_t_plus_one = add_tenor(as_of, '1D', 'MF', 'nyc')

  
    irs = IRS(
        effective=as_of,
        termination=add_tenor(as_of, '10Y', 'MF', 'nyc'),
        notional=10000,
        curves=curves[as_of],
        spec='usd_irs',
    )

    irs_t_plus_one = IRS(
        effective=as_of,
        termination=add_tenor(as_of, '10Y', 'MF', 'nyc'),
        notional=10000,
        fixed_rate=irs.rate(),
        leg2_fixings=fixing.loc[as_of].values[0],
        curves=curves[as_of_t_plus_one],
        spec='usd_irs',
    )
```

Now, I would expect that if I do that every day and cumulate all `irs_t_plus_one.npv()` to match 10Y swap rates times duration plus carry (modulo notional).

Edit: it seems it works better when I use leg2_fixings as a pd.Series of one day rather than passing a float

So we should not pass a float.

## Answer by Attack68 (score 1)

https://quant.stackexchange.com/a/85162

You should always consider swap daily NPV in two stages.

- First is T-1 close to T-0 open. This is transitioning yesterday's curves to account for today's new date, without assuming anything regarding market moves.

- Then you have T-0 open to T-0 close, which accounts for market moves on the day T-0.

If you measure the T-1 close to T-0 open with consistent data in rateslib you should not 'leak' any PnL.

```
curve_T_minus_1_close = Curve(
    nodes={
        dt(2025, 10, 29): 1.0,
        dt(2026, 10, 29): 0.95
    },
    calendar="nyc",
    convention="act360"
)
curve_T_zero_open = curve_T_minus_1_close.translate(dt(2025, 10, 30))

irs = IRS(
    dt(2025, 10, 29),
    "6m",
    spec="usd_irs",
    fixed_rate=2.5,
    notional=1e9,
)
fixing_rate = curve_T_minus_1_close.rate(dt(2025, 10, 29), "1b")
irs_with_fix = IRS(
    dt(2025, 10, 29),
    "6m",
    spec="usd_irs",
    fixed_rate=2.5,
    notional=1e9,
    leg2_fixings=Series(
        index=[dt(2025, 10, 29)],
        data=[fixing_rate],
    )
)

t_minus_1_npv = irs.npv(curves=curve_T_minus_1_close)
print(t_minus_1_npv)
print(irs_with_fix.npv(curves=curve_T_zero_open))
print(t_minus_1_npv*(1+fixing_rate/36000))
```

In the last lines you will find that the NPV as determined by the translated curve exactly matches the projected NPV rolled forward overnight by the rate of the fixing.