Day-0 Discrepancy Audit

Report: max_report.tex, §1 Data Cleaning, Figure 1. Source: forecasting-bias-ibes/max/ scripts and their saved outputs. Audited 2026-09-26.

Figure 1 shows

48.2%

Day-0 share in the plotted histogram (19,370 / 40,198). The data is EPS estimates, not price targets.

Text implies

36.5%

272,741 / 746,626. The denominator only contains Days 0–1, not Days 0–100.

Ground truth

≈12.8%

Day-0 share of 12-month price targets issued 0–100 days after the last earnings call (272,349 / 2,132,602).

Bottom line. The figure and the text come from two different datasets, and both are wrong for the claim they support. Figure 1 is a histogram of the raw I/B/E/S EPS estimate file. 7_forecast_anchor_distance_histogram.py picked it automatically because it was the largest parquet on disk. The "746,626 forecasts within 100 days" in the text is really a file already cut to Days {0, 1}. Rebuilt from the raw price-target file, Day 0 is about 12.8% of forecasts in the 0–100 day window. It is still the busiest single day, but it is not a third or half of the sample.

1What the report claims

The Figure 1 caption and the §1 text say the baseline sample is "12-month I/B/E/S forecasts … published within 100 days post-earnings (746,626 forecasts), of which 272,741 … were published specifically on Day 0". That gives a Day-0 share of 272,741 / 746,626 = 36.5%. The bar at Day 0 in the figure reaches 0.48. These two numbers cannot describe the same data.

2Reverse-engineered funnel

I matched every count in the report to a saved script output or a file row count. The funnel crosses between three different datasets without saying so.

Report step	Count	Where it actually comes from	Status
Raw one-year price targets	1,276,627	Row count of `ibes_nptgsum.parquet`, the summary / consensus file (one row per ticker-month), not analyst-level	consensus level
− recording errors, − unmapped CUSIPs	−1,747 / −77,568	1,276,627 − 1,747 − 77,568 = 1,197,312 = row count of `cleaned_consensus_targets.parquet`. Nothing downstream uses this file.	dead end
"Within the first 100 days" (analyst-level)	746,626	`tight_window_analysis_row_level.parquet`, built from `ibes_nptgdet` (analyst detail). Its own report says `Filter window: days_since_earnings in {0, 1}`. It holds 281,535 Day-0 and 465,091 Day-1 rows and nothing else.	mislabelled
(script 10) drop events with no 4-case match	−26,269 → 720,357	`dataset_filtering_retention_report.txt`. The report skips this step.	omitted
Day 0 (trading-day matched)	272,741	Script 10: 37.86% of 720,357, or 36.5% of 746,626. Either way, this is Day 0 as a share of Days 0–1 only.	count correct
Firm-date median consensus	122,958	`consensus_with_base_prices_diagnostics_sp500.txt`	consistent
S&P 500 universe (size rank 1–500)	31,285	Same diagnostics file	consistent
≥3 analysts on Day 0	28,807 → 13,817	`consensus_with_signals_diagnostics.txt`. The input is 28,807, not 31,285, so it comes from a different run.	run mismatch

3Why the figure shows 48%

The figure file in the report, analyst_days_since_earnings_distribution.png, is written by max/7_forecast_anchor_distance_histogram.py. That script has no fixed input. Its find_full_window_dataset() function:

searches the whole repository for *.parquet files and skips names containing tight, twoday, 2day, filtered or 0day. That rule excludes the real price-target files.
keeps any file that has an earnings-date column (anndats_act) and a forecast-date column (anndats). The EPS detail files also have both.
picks the file with the largest size on disk: ibes_prev_earning_date_with_day_gap.parquet (497 MB, 33.8M rows). This is ibes_ndet (I/B/E/S EPS estimates, fpi 1/2/6/7…, median value 0.68) with extra columns added.

selected = max(candidates, key=lambda f: f.stat().st_size)   # line 43: "the raw untruncated dataset"

In that file, anndats_act is the announcement date of the fiscal period being forecast, so 89.8% of the EPS estimates fall before it (negative gap). Only 40,198 of 33.8M rows land in 0–100, and 19,370 of those are on Day 0: 48.19%. Running the script's selection and binning logic reproduces the plotted 0.482 bar exactly. The hard-coded ax.set_ylim(0, 0.52) was set to fit that bar.

The script also has a guard that rejects the right data:

if pd.isna(max_day) or max_day <= 1:
    raise ValueError("... This file is still a 2-day truncated dataset ...")

The only price-target files available were the {0,1}-truncated ones. The file-name filter and this guard together pushed the script onto the one large file that passed, which was the wrong dataset. The script ran without errors and printed a path that nobody checked.

4Why the text's 36.5% is also wrong

10_filter_0day_after_earnings.py honestly labels its 0–100 window, but the input it filters (tight_window_analysis_row_level.parquet) contains only Days 0 and 1. The "0–100 window" step therefore removes nothing. The 746,626 / 720,357 denominators are "Day 0 + Day 1", and 272,741 / 720,357 = 37.86% is the Day-0 share of a two-day sample. The script's own plot (§6, right) shows this: it has exactly two bars. No script in the repo creates the {0,1} file. Its builder is missing, and max/ is not under git.

5Ground truth: rebuilt 0–100 day sample

Method. I started from ibes_nptgdet.parquet (analyst-level price targets), kept horizon == 12 (2,232,850 rows) and mapped CUSIP→PERMNO with cusip_permno.parquet (2,161,883 rows). I attached each target to the most recent earnings announcement on or before its date for the same ticker (merge_asof, backward), using I/B/E/S actuals (ibes_nact). Then gap = forecast date − earnings date, in calendar days.

Validation against the report's own file. My rebuild gives 726,995 rows at gap ∈ {0,1} (their file: 746,626) and 272,349 at Day 0 (their trading-matched count: 272,741). 97.0% of their rows appear in the rebuild with the same earnings date. Using ibes_ndet for the earnings dates instead gives 12.85%. The result does not depend on that choice.

Window 0–100 days after the last call	Forecasts	Share
Day 0	272,349	12.77%
Day 1	454,646	21.32%
Days 2–100	1,405,607	65.91%
Total in window	2,132,602	100%
Outside window (>100 days, or before the first recorded call)	29,281

If the rebuild is scaled up by the ~2.7% row gap to the report's file, the Day-0 share stays between 12.5% and 12.8%. The report's own downstream number, 272,741 Day-0 forecasts, is essentially right. What is wrong is the denominator: about 2.13M forecasts fall in the 0–100 window, not 746,626.

6The three distributions side by side

Report Figure 1 — Figure 1 as it appears in the report (script 7). The Day-0 bar is 0.482. This is the EPS estimate file.

Script 10 plot — Script 10's own plot of the "0–100 day" window. It has only two bars, which shows the input was already truncated to Days {0,1}.

7Fixes

Replace Figure 1 with a histogram of the rebuilt price-target sample. It should show a Day-0 bar of about 12.8% and a Day-1 bar of about 21.3%.
Rewrite the §1 text: "≈2.13M analyst-level 12-month price targets were published within 100 days after an earnings call, of which 272,741 (≈12.8%) were published on Day 0."
Script 7: remove find_full_window_dataset() and read one explicit, named input. Add assertions that the input is price-target data (for example, a price_target/forecast_price column is present).
Recover or rewrite the missing builder of tight_window_analysis_row_level.parquet. Keep a 0–100 day intermediate file so the Day-0 filter runs on the real window.
Fix the funnel: the 1,276,627 → 1,197,312 steps belong to the consensus summary file. Either drop them or restate the funnel on ibes_nptgdet (2,299,723 raw → 2,232,850 12-month → 2,161,883 mapped). Also report the 26,269-row unmatched-event drop, and rerun so 31,285 and 28,807 agree.

8Reproduce

The scripts are in /root/home/kurby/playground/max_error/repro/. Run them from forecasting-bias-ibes/ with pandas and pyarrow installed.

python repro/select_like_script7.py   # which parquet script 7 picks (EPS file)
python repro/share.py ibes_ndet.parquet max/outputs_max/price_target_pipeline/tight_window_analysis_row_level.parquet
                                     # 48.19% on the EPS file; the tight file only has days 0 and 1
python repro/rebuild.py              # ground-truth 0–100 day price-target sample: Day 0 = 12.77%