US excess saving and consumption
Question: did the pandemic-era stock of "excess saving" get spent, and does US
consumption grow faster in the months when that stock is being drawn down?
Answer, from the data in RESULTS.md: the stock is real and
large — it peaks at about $2.2 trillion in August 2021 — but the month-to-month
link to consumption growth is not there. The correlation between consumption
growth and the drawdown flips sign depending on where the window starts
(+0.62 from 2021-01, +0.07 from 2021-07, -0.22 from 2023-01), and the
positive early reading is almost entirely one month: March 2021, when the third
stimulus payment pushed spending and saving up together. A single window would
have produced a confident but meaningless number in either direction.
Whether the stock is now exhausted depends on an assumption the data cannot
settle: extrapolating the fitted 2015–2019 saving-rate trend says it went
negative in 2024-12, while holding the pre-pandemic average rate flat says
about $1.1 trillion is left. Both are in RESULTS.md; the honest answer is
that the counterfactual, not the data, is doing the work.

Method
- Fit an OLS line to the personal saving rate over 2015-01 … 2020-02 (the
last month before US pandemic transfers). A rate, not a dollar level, so the
counterfactual scales with realised income instead of assuming pre-pandemic
income growth.
- Extrapolate that line and convert it to dollars with each month's realised
disposable income:
counterfactual_saving = rate / 100 * DSPI.
- Monthly excess saving is
(actual_saving − counterfactual_saving) / 12, since
both series are annual-rate dollars while the flow is monthly.
- Cumulate from 2020-03; report the peak, the drawdown and the month it crosses
zero, under both the fitted-trend and the flat-rate counterfactual.
- Correlate consumption growth with (a) the change in the excess stock and
(b) the saving-rate gap, over four window start dates rather than one.
Data
Four monthly BEA series, pulled from FRED's public CSV endpoint — no API key,
no manual download, so anyone can reproduce the numbers:
| Series |
Meaning |
PCE |
Personal consumption expenditures, $bn, SAAR |
PMSAVE |
Personal saving, $bn, SAAR |
PSAVERT |
Personal saving as a % of disposable income |
DSPI |
Disposable personal income, $bn, SAAR |
Downloads are cached in data/ (git-ignored) and reused unless --refresh is passed.
Run it
pip install -e ".[dev]"
python -m pce_savings # uses cached CSVs if present
python -m pce_savings --refresh # re-download from FRED
pytest -q # unit tests, no network
This rewrites RESULTS.md, results.json and figures/. CI runs the tests and
lints on every push, and re-runs the full analysis against live FRED data monthly,
so the repository fails loudly if the pipeline or the source data breaks.
Limitations
- Accounting, not behaviour. Disposable income is spent or saved by
construction, so consumption and saving are mechanically linked; a negative
correlation is partly definitional and none of this identifies causation.
- The counterfactual dominates the level. As shown above, "is it gone?" is
answered by the baseline assumption, not by the data.
- Aggregates hide distribution. Excess saving is concentrated in higher-income
households, so an aggregate stock says little about the households whose
spending is most sensitive to it.
- BEA revises. Personal income and saving are revised for years; re-running
this on a later vintage will move the numbers.
- Monthly Pearson correlation on trending series is a weak test; a proper answer
would use a distributed-lag or error-correction specification.
繁體中文摘要
研究問題:疫情期間累積的「超額儲蓄」是否已經花完,以及美國消費支出是否在超額
儲蓄被消耗的月份成長得更快?
結果:超額儲蓄確實存在且龐大(2021 年 8 月達到約 2.2 兆美元高點),但與消費成長
的月度關聯不成立——相關係數會隨樣本起點翻轉正負(2021-01 起 +0.62、2022-01
起 -0.11、2023-01 起 -0.22),而早期的正相關幾乎全來自 2021 年 3 月一次紓困
發放,當月消費與儲蓄同時跳升。至於「花完了嗎」,取決於反事實假設而非資料本身:
沿用 2015–2019 擬合趨勢會得到 2024 年 12 月即由正轉負,改用疫情前平均儲蓄率則
還剩約 1.1 兆美元。兩種算法都寫在 RESULTS.md。
資料全部來自 FRED 公開 CSV 端點(不需要 API key),python -m pce_savings 就能
從零重現所有數字與圖表;分析的算術部分有單元測試,CI 每月會用最新資料重跑一次。
Repository history
archive/PCE-2023-bea-excel.ipynb is the earlier version of this question: it
read a manually downloaded BEA spreadsheet with hard-coded column offsets, so it
could not be reproduced by anyone else and never reached an answer. It is kept
verbatim as a record and is excluded from lint. An unfinished valuation scratch
file (EPS.py) was deleted rather than published half-working.