PoyrazDemir/market-base-rates

Market base rates (S&P 500)

Reproducible base rates for the US stock market, computed from primary data with plain pandas: long-run returns, what followed high CAPE readings, the cost of missing the best days, what happened to people who waited in cash, drawdown frequency, buying at all-time highs, the real yield on cash, and lump sum vs dollar-cost averaging.

This dataset is the results/ folder of the market-base-rates repository: 33 small CSV tables, every one produced by python run_all.py from the source files listed below. Nothing is typed in by hand. There are no forecasts and no recommendations here, only distributions and counts, with the caveats spelled out in What this does and does not show. A rendered version of all tables is on the project page: https://poyraz-demir.github.io/market-base-rates/.

• Repository (code, source snapshots, tests): https://github.com/poyraz-demir/market-base-rates
• Rendered tables: https://poyraz-demir.github.io/market-base-rates/
• Data vintage: daily series through 2026-09-18 (DATA_END); Shiller monthly data through 2026-09; Damodaran annual data through 2025; FRED CPI through 2026-08. Tables generated 2026-09-21.

Loading

Each table is its own config (the tables have different columns, so they cannot share one config). Config names equal the file names without .csv; the two-digit prefix is the topic group. The default config is 02_forward_by_bucket.

from datasets import load_dataset

# one table
ds = load_dataset("PoyrazDemir/market-base-rates", "05_episodes_ge20", split="train")
ds.to_pandas()

# every table of one topic group
from datasets import get_dataset_config_names
names = [n for n in get_dataset_config_names("PoyrazDemir/market-base-rates") if n.startswith("02_")]
tables = {n: load_dataset("PoyrazDemir/market-base-rates", n, split="train").to_pandas() for n in names}

Or with pandas directly:

import pandas as pd
df = pd.read_csv("hf://datasets/PoyrazDemir/market-base-rates/data/02_forward_by_bucket.csv")

Conventions: returns, shares and CAGRs are decimals (0.1096 = 10.96 %); multiples are plain numbers (17.55 = 17.55x); durations are calendar days unless the column says otherwise; dates are ISO strings (YYYY-MM-DD, or YYYY-MM for months, or a bare year). Values in 07_rates_snapshot.csv are in percent as FRED publishes them. Empty cells mean "not applicable / never triggered".

Files

File	Rows	What it contains	Columns
01. Long-run returns (scripts/01_long_run_returns.py)		S&P 500 CAGR over 10/20/30/50/98 years, nominal and real; every rolling 10-year window; other asset classes (Damodaran annual data, 1928-2025).
01_annual_distribution.csv	1	Calendar-year S&P 500 total-return distribution 1928-2025: number and share of negative years, worst and best year, arithmetic and geometric mean.	years, negative_years, share_negative, worst_year, worst_return, best_year, best_return, arithmetic_mean, geometric_mean
01_asset_classes.csv	6	Nominal CAGR to end-2025 over 10/20/30/50/98 years for the S&P 500 (TR), gold, Baa corporate bonds, US home prices, 10-year T-bond and 3-month T-bill, plus real 30-year CAGR.	asset, 10y, 20y, 30y, 50y, 98y, real_30y
01_cagr_windows.csv	5	S&P 500 nominal, real and inflation CAGR over the 10/20/30/50/98-year windows ending 2025, and growth of 100 (nominal).	window, years, nominal_cagr, real_cagr, inflation_cagr, growth_of_100
01_rolling_10y.csv	89	Every rolling 10-year window 1928-1937 ... 2016-2025 (89 windows): nominal and real CAGR.	start, end, cagr_10y_nominal, cagr_10y_real
01_rolling_10y_summary.csv	1	Summary of the rolling 10-year windows: median, min and max with their windows, share of windows negative nominal and real.	windows, median, min, min_window, max, max_window, share_negative, share_negative_real
02. CAPE and forward returns (scripts/02_cape_forward_returns.py)		Forward 1/3/5/10-year real and nominal returns after months with CAPE 25-30, >30, >35, >40; count of independent episodes behind each bucket; 1929 & 1997-2002 vs 2017-2021 (Shiller monthly data, 1871-2026).
02_cape_gt40_episodes.csv	2	Every contiguous run of months with CAPE > 40 (start, end, length).	run_start, run_end, months
02_cape_now.csv	1	Latest CAPE reading vs the full-history median and mean, its percentile, months above it, and the all-time peak.	month, cape, cape_median_all, cape_mean_all, percentile, months_above_current, months_with_cape, peak_cape, peak_month
02_forward_by_bucket.csv	40	Forward 1/3/5/10-year annualised return after months in CAPE buckets all / 25-30 / >30 / >35 / >40, real (kind=real) and nominal (kind=nom); overlapping monthly windows, so n counts windows, not independent observations.	bucket, months_in_bucket, horizon_years, kind, n, median, mean, worst, best, share_negative, p5, p25, p75, p95
02_independent_episodes.csv	7	Contiguous runs of months in each CAPE bucket that already have a full 10-year outcome, with the median 10-year real and nominal return of each run.	bucket, run_start, run_end, months, median_real_10y, median_nom_10y
02_regime_split.csv	12	CAPE > 30 entries split into two regimes (1929 & 1997-2002 vs 2017-2021): forward 1/3/5-year statistics, real and nominal.	regime, horizon_years, kind, n, median, mean, worst, best, share_negative, p5, p25, p75, p95
03. Missing the best days (scripts/03_missing_best_days.py)		CAGR without the 10/20/30/50/100 best days, without the worst days, without both; where the best days sit relative to drawdowns and worst days (S&P 500 price 1928-, total return 1988-, SPY 1993-).
03_best_days_by_year.csv	41	How many of the 50 best days of each series fall in each calendar year.	series, year, best_days
03_best_days_list.csv	300	The 50 best and 50 worst single trading days of each of the three series (300 rows): date, daily return and the drawdown from the running high at that close.	series, rank, kind, date, return, drawdown_at_close
03_best_days_location.csv	3	How many of the 50 best days sit inside a >20 % / >10 % drawdown and within 5 trading days of one of the 50 worst days.	series, best_days, in_drawdown_gt20, in_drawdown_gt10, within_5_days_of_a_worst_day
03_missing_days.csv	36	Buy and hold vs without the 10/20/30/50/100 best days, without the 10/20/30/50/100 worst days, and without both: multiple, CAGR and growth of 10,000 over the same calendar span, for three series.	series, start, end, years, trading_days, scenario, multiple, cagr, growth_of_10k
04. Waiting in cash (scripts/04_waiting_in_cash.py)		Moving to 3-month T-bills on 1996-12-05, 2014-01-02, 2018-01-02, 2021-01-04, 2025-01-02 vs investing the same day; was cash ever ahead; waiting for a 10/20/30 % dip before entering.
04_wait_for_dip.csv	5	Stay in T-bills until the index is 10/20/30 % below its running high since the start date, then go all in: final multiple, entry date (blank = never triggered) and difference vs investing immediately.	start, invest_immediately, wait_for_10pct, wait_for_10pct_entry, wait_for_10pct_vs_immediate, wait_for_20pct, wait_for_20pct_entry, wait_for_20pct_vs_immediate, wait_for_30pct, wait_for_30pct_entry, wait_for_30pct_vs_immediate
04_waiting_nominal.csv	5	Moved to 3-month T-bills on the start date vs invested the same day in the S&P 500 (TR), nominal: years, multiples, CAGR, outcome of 100,000, cash shortfall.	start, reason, years, stocks_multiple, stocks_cagr, cash_multiple, cash_cagr, stocks_100k, cash_100k, cash_shortfall
04_waiting_real.csv	5	Same comparison CPI-deflated: inflation multiple, real multiples and real CAGR for stocks and cash.	start, inflation_multiple, stocks_real_multiple, stocks_real_cagr, cash_real_multiple, cash_real_cagr
04_was_cash_ever_ahead.csv	5	Lowest stocks-to-cash ratio and its date, last day cash was ahead, share of trading days cash was ahead.	start, min_stocks_to_cash, min_date, last_day_cash_ahead, share_of_days_cash_ahead
05. Drawdowns (scripts/05_drawdowns.py)		Frequency of 5/10/20/30/50 % declines in the price index, peak-to-trough and recovery durations, every >=20 % episode, intra-year drawdowns by year.
05_drawdown_frequency.csv	5	For 5/10/20/30/50 % thresholds: number of episodes, years per episode, median and worst depth, median peak-to-trough, trough-to-recovery and full-cycle durations (calendar days), mean and max full cycle in years, open episodes.	threshold, episodes, years_per_episode, median_depth, worst_depth, median_peak_to_trough_days, median_trough_to_recovery_days, median_full_cycle_days, mean_full_cycle_years, max_full_cycle_years, open_episodes
05_episodes_ge20.csv	12	Every decline of 20 % or more in the S&P 500 price index since 1928: peak, trough, depth, days to trough, recovery date, full cycle in years.	peak, trough, depth, peak_to_trough_days, recovery, full_cycle_years
05_intra_year.csv	2	Intra-year maximum drawdown statistics for 1928-2025 and 1980-2025: mean and median, positive calendar years, share of years with a drawdown of at least 5/10/15/20 %.	period, years, mean_intra_year_drawdown, median_intra_year_drawdown, positive_calendar_years, share_years_dd_ge_5pct, share_years_dd_ge_10pct, share_years_dd_ge_15pct, share_years_dd_ge_20pct
05_intra_year_by_year.csv	100	Per calendar year 1927-2026: intra-year maximum drawdown and calendar-year return of the price index (2026 is partial, through DATA_END).	year, intra_year_max_drawdown, calendar_return
05_time_below_high.csv	4	Share of trading days spent 10/20/30 % or more below the running high, for the full history and the last 50/30/20 years.	period, days, below_high_by_10pct_or_more, below_high_by_20pct_or_more, below_high_by_30pct_or_more
06. Buying at all-time highs (scripts/06_near_ath.py)		Share of trading days at or near the all-time high; forward 1/3/5-year returns after buying at a high vs any day vs 10-20 % below the high.
06_forward_after_entry.csv	36	Forward 1/3/5-year annualised return after entering on a random day, a new all-time-high day, within 1 % / 5 % of the high, or 10 %+ / 20 %+ below it; S&P 500 total return (1988-) and price index (1928-).	series, horizon_years, entry, n, median, mean, worst, best, share_negative, p5, p95
06_share_near_high.csv	4	Share of trading days that are a new all-time high, within 1/2/5/10 % of it, or more than 20 % below it, by period.	period, days, new_high_days, share_new_high, within_1pct, within_2pct, within_5pct, within_10pct, below_high_gt20pct
07. Real yield on cash (scripts/07_cash_real_yield.py)		Real 3-month T-bill yield now vs 2021 and by period since 1982; share of months with a negative real yield; worst months; calendar-year cash return vs CPI; snapshot of other FRED rates.
07_calendar_years.csv	44	Calendar-year return of a 3-month T-bill cash index vs December-to-December CPI, and the real return, 1982-2025.	year, cash_return, cpi_dec_to_dec, real_return
07_current.csv	4	Real 3-month T-bill yield now (latest daily and latest monthly basis) vs the 2021 average and the 2010-2019 average.	as_of, basis, nominal, cpi_yoy, real
07_negative_months_by_decade.csv	7	Share of months with a negative real cash yield and the mean real yield, by period.	period, months, share_months_real_negative, mean_real
07_rates_snapshot.csv	5	Last value in the snapshot of FRED DGS3MO, DGS30, T10YIE, ECBDFR and ECBMRRFR (percent, not decimals).	series, description, date, value
07_real_cash_by_period.csv	9	Monthly real 3-month T-bill yield by period 1982-2026: mean nominal, mean CPI y/y, mean/min/max real, share of months negative.	period, months, mean_nominal, mean_cpi_yoy, mean_real, min_real, max_real, share_months_real_negative
07_worst_months.csv	10	The ten months with the most negative real cash yield (monthly average 3-month yield minus CPI y/y).	month, nominal, cpi_yoy, real
08. Lump sum vs dollar-cost averaging (scripts/08_lump_sum_vs_dca.py)		Lump sum vs 3/6/12-month DCA over 453 rolling 12-month windows from 1988 (S&P 500 total return), with and without T-bill interest on the uninvested cash.
08_lump_sum_vs_dca.csv	6	Lump sum vs 3/6/12-month DCA, wealth after 12 months over 453 rolling windows: share of windows lump sum wins, median and mean gap, 5th/50th/95th percentile of each, share of windows DCA beats staying in cash.	dca_months, uninvested_cash, windows, ls_beats_dca, median_ls_minus_dca, mean_ls_minus_dca, ls_p5, dca_p5, ls_median, dca_median, ls_p95, dca_p95, dca_beats_all_cash
08_windows.csv	453	Every 12-month window (monthly entries 1988-01 ... 2025-09): wealth multiple after 12 months for lump sum, all cash in T-bills, and DCA over 3/6/12 months with the uninvested part in T-bills or at 0 %.	start, end, lump_sum, all_cash_tbill, dca_3m_tbill, dca_3m_zero, dca_6m_tbill, dca_6m_zero, dca_12m_tbill, dca_12m_zero

Method, briefly

• Sources are read as downloaded, no manual edits (scripts/common.py: load_shiller, load_damodaran, load_fred, load_daily).
• DATA_END = 2026-09-18. Every daily series is truncated at that date, so a later download of the same Yahoo Finance symbols reproduces these tables exactly (history for past dates is stable; only the tail grows).
• Cash index. 3-month T-bill yields (FRED DGS3MO, daily from 1981-09) are compounded day by day into a cash index with continuous reinvestment and no taxes; missing days (holidays) are dropped.
• Overlapping windows. All forward-return tables (groups 02, 06, 08) use monthly or daily start dates. Consecutive windows share almost their whole path, so n counts windows, not independent observations. 02_independent_episodes.csv lists the contiguous runs behind each CAPE bucket: CAPE > 30 ten-year statistics rest on four runs, CAPE > 35 on two runs inside the dot-com bubble, and CAPE > 40 is a single episode (1999-01 to 2000-09) counted 21 times. The second CAPE > 40 run (2026-05 to 2026-09) has no forward outcome yet.
• Drawdowns are computed on the S&P 500 price index (^GSPC, no dividends) from 1927-12-30; an episode starts at a running high and ends when that high is regained.
• Missing the best days removes the top-k daily returns from the series and compounds the rest over the same calendar span; the worst-days and both-removed scenarios are computed the same way.
• Real values deflate by CPI: FRED CPIAUCSL for the daily-series scripts, Shiller's CPI column for the CAPE script, Damodaran's inflation column for the annual script.
• Lump sum vs DCA spreads the purchase evenly over 3/6/12 monthly entries and compares wealth after 12 months, with the uninvested part earning either the 3-month T-bill rate or 0 %.

Sources and attribution

Only derived tables are included here. The primary series are not redistributed in this dataset; the first three are committed as unmodified snapshots in the repository (data/raw/), the Yahoo Finance series are downloaded by fetch.py. Download dates are when the snapshot was taken.

Source	Content	Used by	Downloaded	Terms
Robert J. Shiller, Yale University, Irrational Exuberance data (ie_data.xls), https://shillerdata.com/	Monthly S&P composite price, dividends, earnings, CPI, GS10, real total-return index, CAPE; 1871-01 to 2026-09	02	2026-09-20 (file last saved by the author 2026-09-02; September price = 1 September close; latest CPI months are the author's estimates)	Published for public use, customarily reused with attribution. Cite: Shiller, R. J., Irrational Exuberance, Princeton University Press (2000, 2005, 2015), data updated at shillerdata.com.
Aswath Damodaran, NYU Stern, "Historical Returns on Stocks, Bonds and Bills: 1928-Current" (histretSP.xlsx), https://pages.stern.nyu.edu/~adamodar/pc/datasets/histretSP.xlsx	Annual returns 1928-2025 on the S&P 500, 3-month T-bills, 10-year T-bonds, Baa corporate bonds, US home prices, gold, US CPI; nominal and real	01	2026-09-19 (sheet header "Date updated: 2026-01-01"; file last saved 2026-08-24)	Freely available for use with attribution. Cite: Damodaran, A., "Historical Returns on Stocks, Bonds and Bills: 1928-Current", NYU Stern.
FRED, Federal Reserve Bank of St. Louis, https://fred.stlouisfed.org/: DGS3MO, DGS30, T10YIE, ECBDFR, ECBMRRFR, CPIAUCSL, CPIAUCNS	Daily Treasury constant-maturity yields (Board of Governors, H.15), 10-year breakeven inflation (FRB St. Louis), ECB policy rates (ECB), CPI-U SA and NSA (US BLS)	04, 07, 08	2026-09-19 (yields through 2026-09-17/18, CPI through 2026-08)	Redistribution permitted with attribution: "Source: FRED, Federal Reserve Bank of St. Louis; original source as listed." See https://fred.stlouisfed.org/legal/.
Yahoo Finance daily series ^GSPC (close, 1927-12-30-), ^SP500TR (close, 1988-01-04-), SPY (adjusted close, 1993-01-29-)	S&P 500 price index, S&P 500 total-return index, SPY ETF adjusted close (dividends reinvested, after the fund's expenses)	03, 04, 05, 06, 08	2026-09-20, last observation 2026-09-18	Licensed for personal, non-commercial use; may not be redistributed. The raw series are therefore not included anywhere; only these aggregated results are.

The ^GSPC series was cross-checked against J.P. Morgan's Guide to the Markets (average intra-year decline 1980-2026 YTD: JPM -14.2 %, this series -14.1 %; positive calendar years 35 of 46 in both).

Reproduce

git clone https://github.com/poyraz-demir/market-base-rates.git && cd market-base-rates
python -m venv .venv && . .venv/bin/activate
pip install pandas openpyxl xlrd          # or: pip install -e ".[dev]"
python fetch.py                           # downloads the Yahoo Finance daily series into data/cache/
python run_all.py                         # writes results/*.csv, results/SUMMARY.md and refreshes the README
pytest                                    # smoke tests + a few known numbers with tolerances

Python 3.10+. The files in this dataset should be byte-identical to results/*.csv after run_all.py on the same source snapshots. Each topic script also runs on its own (python scripts/05_drawdowns.py).

What this does and does not show

• Overlapping windows. All forward-return tables (groups 02, 06, 08) use monthly or daily start dates, so consecutive windows share almost their whole path. n counts windows, not independent observations. Group 02 lists the contiguous runs behind each CAPE bucket: the CAPE > 30 ten-year statistics come from four runs, the CAPE > 35 ones from two runs inside the dot-com bubble, and CAPE > 40 is a single episode (1999-2000) counted 21 times. Non-overlapping variants (e.g. January entries only) leave too few observations to describe a distribution, which is why the overlapping tables are shown together with the episode count.
• Two regimes, one bucket. CAPE > 30 entries in 1929 and 1997-2002 were followed by negative real returns; CAPE > 30 entries in 2017-2021 were followed by strongly positive ones. The pooled median describes neither. 58 % of CAPE > 30 months (entries from mid-2017 on) do not have a ten-year outcome yet.
• Daily total return only from 1988. Groups 03, 05 and 06 use the price-only ^GSPC series before 1988 (no dividends), which understates absolute returns by roughly 3-4 percentage points a year but barely affects comparisons within one series. The 1929-1954 "25-year recovery" is nominal, price-only; with dividends reinvested and 1930s deflation the real recovery came earlier.
• Cash series only from September 1981 (start of FRED DGS3MO), so the real-cash-yield history (07) starts in 1982 and the waiting-in-cash cases (04) cannot go back further. TB3MS/DTB3 were not obtained.
• Missing best days vs missing worst days are symmetric calculations; the asymmetry people quote exists because best and worst days cluster in the same crashes (27-31 of the 50 best days sit within 5 trading days of one of the 50 worst, depending on the series).
• No taxes, no fees, no slippage anywhere except the SPY series (which is net of the fund's expense ratio).
• Recent CPI in Shiller's file is estimated by the author for the last few months; the September price is the 1 September close. FRED CPI runs through August 2026.
• Lump sum vs DCA (08) follows the design of Vanguard's 2023 paper but on a different index and period (S&P 500 total return from 1988 vs Russell 3000 and MSCI World from 1976-79), so hit ratios are comparable in spirit, not identical.
• Not obtained, so not reproduced: Vanguard 2012 "Dollar-Cost Averaging Just Means Taking Risk Later" (links dead), DALBAR QAIB and Morningstar "Mind the Gap" (paywalled / blocked), Hartford Funds / Ned Davis "missing the best days" tables (403), S&P Dow Jones Indices' own 100-year total-return series (403).
• Single-row tables. Several files (01_annual_distribution, 01_rolling_10y_summary, 02_cape_now) are one-row summaries kept as separate tables so that every number on the project page has a file behind it.

Layout of this dataset

README.md      this card
LICENSE        MIT (code and these derived tables; sources keep their own terms)
SUMMARY.md     all 33 tables rendered as markdown (copy of results/SUMMARY.md)
verify.py      sha256 check of data/*.csv and SUMMARY.md against a local clone's results/
data/          33 CSV tables, byte-identical to results/*.csv in the repository

License and citation

The tables and the code that produces them are MIT-licensed (see LICENSE). Each primary source keeps its own terms, listed above. This dataset is not investment advice.

@misc{market_base_rates_2026,
  title        = {Market base rates (S\&P 500): reproducible distributions and counts},
  author       = {Demir, Poyraz},
  year         = {2026},
  howpublished = {\url{https://github.com/poyraz-demir/market-base-rates}},
  note         = {Dataset on Hugging Face: PoyrazDemir/market-base-rates}
}