Methodology

Forecasting inflation and earnings

The Triple Lock uses two readings each year: CPI inflation in September, and growth in average weekly earnings over May to July. Both are published by the ONS. The model forecasts them from a monthly statistical model fitted to 303 months of ONS data, from February 2001 to April 2026: the CPI index (series D7BT) and the level of average weekly earnings, total pay (series KAB9).

The model works in month-on-month changes rather than annual rates. Consecutive annual rates share eleven months of data, so a model fitted to them mistakes that overlap for persistence and produces forecasts that are too smooth. Month-on-month changes do not overlap. Each simulated path is turned back into a September CPI rate and a May to July earnings rate by compounding twelve months at a time, the same construction the ONS uses. Applied to the historical data, this reproduces the published September CPI rate exactly, and the published May to July earnings rate to within 0.13 percentage points.

Monthly inflation depends on its own value one, six and twelve months earlier. The six- and twelve-month terms pick up seasonal patterns, since the CPI index is not seasonally adjusted. Monthly earnings growth depends on its own value in the previous month and on inflation in the previous one and two months. Earnings do not feed back into inflation. This link matters for the Triple Lock: an inflation shock that sets one year's uprating also pushes up the earnings reading that can set the next.

Each forecast adds random shocks, drawn from a normal distribution whose size is estimated from the historical data. Drawing 10,000 sets of shocks gives 10,000 monthly paths for inflation and earnings from May 2026 to 2047. Every uprating rule on this site is applied to the same 10,000 paths, so the differences between rules come from the rules themselves, not from different draws.

Normal shocks understate how often large inflation surprises happen. In the simulations, September CPI has a standard deviation of 1.78 percentage points, against 2.24 in the historical data, so an episode like 2022–23 comes up less often than history suggests. The Triple Lock pays the highest of three measures, so it gains most from large shocks. The model therefore probably understates both the cost of the Triple Lock and the savings from replacing it.

From uprating to spending

Spending in each year is the number of pensioners on each type of State Pension, multiplied by the full weekly rate for that type: £241.30 for the New State Pension and £184.90 for the Basic State Pension in 2026/27, uprated each year by the rule being tested. Using full rates ignores partial entitlements, and the additional State Pension, which rises with CPI rather than the Triple Lock. The savings scale with this spending base, so they carry the same approximation.

The two types matter because they pay different amounts. The Basic State Pension covers people who reached State Pension age before April 2016, and the New State Pension covers everyone since. Nobody new joins the Basic State Pension, so the model projects it as a closed group that shrinks with mortality. The New State Pension gains everyone reaching State Pension age, with take-up of 96.8% calibrated to DWP's November 2025 caseload. State Pension age rises from 66 to 67 between 2026 and 2028, as set by the Pensions Act 2014, which slows the inflow for two years. The New State Pension overtakes the Basic State Pension in 2029, and by 2047 it covers 15.1 million people against 0.4 million.

Measuring the savings

The saving in each year is spending under the Triple Lock minus spending under the alternative, on the same forecast path, averaged across all 10,000 paths. Charts show this in cash terms. Totals are also given as a present value in 2026 prices: each year's saving is deflated by that path's own CPI and then discounted at 3.5% a year, the real discount rate in HM Treasury's Green Book. Where ranges are shown, they cover the middle 80% of paths (the 10th to 90th percentile), or the middle 95% for the uprating trajectory.

Income tax on the State Pension

A higher State Pension raises more income tax, so part of each saving comes back to the Exchequer anyway. The net figures subtract that tax. The model takes each pensioner's own taxable income from the Family Resources Survey 2024/25, grows the State Pension part with the uprating rule and the rest with earnings, and applies income tax at 20%, 40% and 45%. Thresholds stay frozen until 2030/31, as announced at Budget 2025, and rise with CPI after that. About 3.9% of State Pension spending goes to people living abroad (DWP figures for 2024/25), who generally pay tax in their country of residence, so that share is left out of the tax calculation.

As a check, the survey puts 68.4% of people over State Pension age above the Personal Allowance. HMRC's count of 9.08 million taxpayers over State Pension age in 2025/26 implies about 67.8%. The tax comes to between 17% and 20% of the gross saving, depending on the alternative, and rises over time as frozen thresholds bring more pensioners into tax.

How each alternative is modelled

Each alternative is a rule that turns the year's CPI and earnings readings, and for some rules past readings, into that year's uprating. Each rule applies from the April 2030 uprating. Rules that look back, such as the CBP Lookback Reform, count the Triple Lock upratings of 2027 to 2029 as part of their history.

What the model leaves out

The baseline is the model's own projection of the Triple Lock, not the OBR's forecast, so savings will not match a costing scored against the OBR's Economic and Fiscal Outlook. Inflation and earnings do not respond to pension policy, and nobody changes their behaviour. The model does not include Pension Credit, whose spending would rise for some pensioners if the State Pension grew more slowly, offsetting part of the saving.

The forecasts also have four known weaknesses, which we are working on. First, inflation settles too high. Nothing anchors it to the Bank of England's 2% target: over 2032–47 it averages 2.9% a year, a level set by the 2001–26 data. This mainly raises the cash figures; the present values are deflated by each path's own inflation, so they are less affected.

Second, earnings growth is too volatile. Year-to-year variation in simulated May to July earnings growth is 2.6 percentage points, against 2.2 in the historical data. Part of this comes from the pandemic: the furlough-distorted months from March 2020 to September 2021 account for about a quarter of the unexplained variation in the earnings equation. Because the Triple Lock pays the highest of three measures, extra volatility in earnings raises its cost, and so the savings from replacing it. This works in the opposite direction to the thin-tailed inflation shocks described above, so the net effect is unclear.

Third, the ranges are too narrow. Every simulation uses the same estimated coefficients, so the 10th to 90th percentile ranges reflect uncertainty about future shocks but not about the model itself. Fourth, the data end in April 2026. The May to July 2026 earnings figure, one of the three readings for the April 2027 uprating, has since been published but is still simulated here.

Separately, past upratings are rebuilt from the latest ONS data, and earnings figures have been revised since each uprating was set. Some years therefore differ from the rise actually paid by up to 0.3 percentage points (8.4% against 8.5% in 2024, and 4.4% against 4.1% in 2025). This affects the charts of past upratings on the first page, and the rules that look back at past upratings, such as the CBP Lookback Reform.

Together with the full-rate spending base described above, these limitations set the precision of the estimates: they are a guide to the relative cost of each rule, and to the rough scale of the savings.

Data sources