1. Why can inflation fall while prices keep rising?
Inflation is a rate of change in a specified price index, not the index level itself. A lower positive twelve-month rate means prices increased less over that comparison window; it does not mean the latest price level fell. When a large earlier monthly increase leaves the twelve-month window, the annual comparison can slow even if the latest month still rises. A regime rule using annual inflation acceleration therefore answers a different question from one using the latest monthly change. Identify the numerator, comparison month and rate horizon before assigning a directional category, and retain the underlying levels so readers can inspect the distinction.
2. Seasonal adjustment changes the measurement object
Seasonal adjustment seeks to remove recurring calendar-related patterns so adjacent-period comparisons are more interpretable. An unadjusted annual comparison and an adjusted monthly comparison can both be useful, but they are not interchangeable observations. BLS maintains specific CPI seasonal-adjustment documentation and calculation guidance. Cite the precise series and treatment rather than describing a generic inflation number. Avoid splicing adjusted and unadjusted levels into one return formula or assuming every component uses the same adjustment method. A label archive should identify the source series, adjustment status and transformation independently, because identical-looking percentages can represent quite different measurement contracts and comparison horizons.
3. Work the comparison-base example from levels
Use invented index levels for two adjacent annual comparisons. The current sequence rises from one hundred ten to one hundred eleven, while the corresponding year-earlier levels rise from one hundred to one hundred two. The annual rates are ten percent and approximately 8.8235 percent, a decline of about 1.1765 percentage points. Yet the current one-month index change is positive 0.9091 percent. The table makes the different denominators visible. This is an arithmetic illustration of comparison-base effects, not an actual CPI release or a decomposition of an observed inflation episode. No seasonal factors, empirical forecast or official estimate are supplied by these constructed numbers.
4. Annualization is not a longer observed history
Annualizing a short-period change expresses what compounded growth would be if that rate repeated through a year. It does not create the missing future observations or establish that repetition is likely. For quarterly changes, BEA explains why it publishes annual rates and how compounding relates the two conventions. A classifier must not compare an annualized quarterly change directly with an unannualized monthly change merely because both are percentages. Store the original rate, annualization formula and assumed periods per year. Keep annualized adjacent-period change distinct from an observed twelve-month change; both carry annual language while relying on different underlying time intervals.
5. Revision risk belongs to the transformed feature too
A feature computed from index levels inherits the vintage of every level used, not just the newest one. If either comparison endpoint is revised, the change rate and its acceleration can move even when the latest displayed level barely changes. Preserve the exact endpoint identities used at the cutoff and recompute revisions as a separate view. A current download containing adjusted historical data does not establish what a contemporaneous analyst saw. FRED's real-time-period documentation supports this distinction. Where the source does not provide adequate vintage or intraday availability evidence, state that limitation instead of assigning a historically precise label from the latest series.
6. Do not let overlapping horizons multiply evidence
Successive twelve-month changes share most of their underlying monthly observations. A long run of declining annual rates can therefore reflect repeated comparison of heavily overlapping windows rather than many independent pieces of evidence. Monthly noise and annual smoothing answer different descriptive questions, neither automatically superior. Define whether the rule measures level, rate, acceleration or surprise relative to an independently timestamped expectation. Combining these dimensions after seeing which produces favorable classifications creates another selection problem. A transparent sensitivity exercise changes one transformation at a time and explains disagreements through the underlying endpoints rather than presenting the selected horizon as uniquely correct.
7. Questions before interpreting a price-state label
Ask whether inflation down means a lower positive rate, a negative rate or a rate below a benchmark. Which index and seasonal treatment are used? Which base observation left the rolling window? Were both endpoint vintages available at the decision? Would the label change under a different declared horizon while using the same eligible data? Common failures include confusing percentage points with percent changes, mixing annualized and observed annual rates, and equating disinflation with falling prices. Report the level movement and the rate movement separately. This makes a descriptive regime statement useful without implying a universal asset preference or future return.
Inspect the comparison
| Synthetic comparison | Current index | Year-earlier index | Twelve-month change |
|---|---|---|---|
| First month | 110 | 100 | 10.0000% |
| Next month | 111 | 102 | 8.8235% |
| Current adjacent-month change | 111 / 110 - 1 | Not a year-earlier comparison | +0.9091% |
Primary references
References support the stated methods and data conventions. Original fictional examples are not observations from the cited sources.