Why is a moving average across a stock split simply wrong?

Because the column it averages contains two different units, and the mean of a mixture of units is not a quantity. This is the adjustment problem from the Price and Trading Data course — and from adjusted-prices in Markets Foundations — wearing a different hat. There it made a chart look strange. Here it produces a plausible, well-formatted, entirely fictional number.

The event

/splits/AAPL.US returns two entries for the 2014–2020 period, one of them:

date: 2020-08-31    split: 4.000000/1.000000

Four new shares for each old one, effective 2020-08-31. Now the raw close column across that boundary, straight from /eod:

2020-08-27   500.04
2020-08-28   499.23
2020-08-31   129.04
2020-09-01   134.18
2020-09-02   131.40

Three averages of the same week

On raw close:

(500.04 + 499.23 + 129.04 + 134.18 + 131.40) ÷ 5 = 1,393.89 ÷ 5 = 278.778

The stock closed at 131.40 that day. The "five-day average price" came out at 278.78 — 2.12 times the price. Nothing was miscalculated. Before 2020-08-31 a row means one old share; from that date it means one new share, and there are four of them per old one.

On split-adjusted prices — divide the two pre-split closes by 4, giving 125.0100 and 124.8075:

(125.0100 + 124.8075 + 129.04 + 134.18 + 131.40) ÷ 5 = 644.4375 ÷ 5 = 128.8875

On EODHD's adjusted_close, which removes dividends as well as splits — 121.2564, 121.06, 125.1654, 130.1511, 127.4545, as fetched on 2026-07-28:

÷ 5 = 125.0175

Three answers: 278.778, 128.8875, 125.0175. The first is broken; the unadjusted mean is 2.16 times the split-adjusted one, and it stays broken for one bar fewer than the window is long — forty-nine days, for a fifty-day average, because the fiftieth post-split bar is the first whose whole window sits after the split.

The other two are both correct, and they are not the same

Split-only adjustment answers "what would this bar look like at today's share count". Full adjustment answers "what did a holder who reinvested dividends actually experience". For 2020-08-28 those are 124.8075 and, on that same pull, 121.06 — a gap of 3.7475, or 3.00% of the split-only figure, being the dividends paid since. Only the first of those two numbers is fixed: the third average above is the one that will differ when you run this, and by the same factor on every row.

So an indicator computed on adjusted_close is a total-return indicator. It will not line up with a chart drawn on split-adjusted prices, and the divergence grows the further back you look and the more the instrument pays out. Neither series is wrong; they answer different questions, and mixing them within one study is what goes wrong.

What the endpoint gives you

/technical/{ticker} exposes splitadjusted_only, an integer 0 or 1 defaulting to 0, which for selected functions restricts adjustment to splits alone. There is also a splitadjusted function that simply returns split-adjusted OHLCV, so you can inspect the series an indicator is running over rather than infer it.

EODHD documents the technical endpoint as computing on split-adjusted OHLC. "Documented" and "true for the function you called with the parameters you sent" are different claims, and the answer turns out to be fully adjusted, dividends included — which is a stronger statement than the documentation makes.

Here is how you can tell, and it is worth doing yourself. On 2026-07-27 AAPL.US had close and adjusted_close both equal to 336.91, so on that day every possible column choice gave the same answer and the test revealed nothing. After the next ex-dividend the two columns separated, and re-running SMA(2) over the identical two days returned the mean of the adjusted pair, not the raw one. A split-only series would not have moved; a dividend moved it. That is the proof, and it only became available because time passed.

Try it now

1. Pull /eod for AAPL.US from 2020-08-26 to 2020-09-02 and reproduce all three averages above. Getting 278.778 with your own code is the moment the problem stops being theoretical.

2. Here is /technical/AAPL.US?function=splitadjusted over 26 August to 2 September 2020. Compare its closes with the raw close column above, row by row, and with adjusted_close. You are looking directly at the difference the parameter makes: find the rows where it changed nothing, and check that its 28 August value is the 124.8075 in the lesson.

3. Here are nine raw closes across the same split, 25 August to 4 September 2020. Compute a 5-day SMA on them for every day it can be computed, and count how many consecutive values mix pre-split and post-split rows. Then say how many it would be for a 50-day SMA. It heals on its own, which is why nobody notices.