Sorry if this is a dumb question, I'm a new user and couldn't find the answer anywhere. To prevent look-ahead bias, I assume the original (unrestated) data is used when the backtest date is prior to the restatement date, right?
Not necessarily. FactSet revises its past data and Portfolio123 hasn't been able to guard the old data from these revisions. Compustat data, if you can afford it, is more point-in-time.
I thought I read somewhere that they capture 3 or 4 revisions. Anything from 2020 onward when they started using FactSet. So it's like we have an improved and more pit version of FactSet.
I could be wrong though...probably am...
1 Like
Ok, well just to keep things simple for now, let's assume the backtest date is prior to 2020. I'm just trying to figure out whether P123 uses the restatement date, if it exists, the way I described.
Well, I know FactSet updates their database when they find an error, but the case I'm talking about here is when the issuer files officially restated financial statements with the SEC. For example, there might be an Original June'24 10Q filed on August 15th, 2024, and a Restated June'24 10Q filed on November 28th, 2024. If the backtest date is in the range August 16th - November 28th, only the Original June'24 10Q would be known to market participants. Both versions of the June'24 10Q are available on FactSet.
Does P123 handle this properly?
I heard the same "rumor" implying that one should pay closer attention to data post 2019 because it's more PIT
We handle restatements as follows. If it's interim, we ignore it. If the restatement is within 300 days of the period end date, we ignore it.
For restatements that are not ignored, we handle them as follows. FactSet gives us two dates for restatements, and we use the later one. So between the period end date and the later of the restatement dates, you'll see the older data, and after the later restatement date, you'll see the newer data. If the restatement ends up being rewritten by a subsequent restatement, only the latter will be available and the effective date will change.
1 Like
We handle restatements as follows. If it's interim, we ignore it. If the restatement is within 300 days of the period end date, we ignore it.
Can you explain the logic? Why ignore all interim restatements and annual ones within 300 days of the period end date?
For restatements that are not ignored, we handle them as follows. FactSet gives us two dates for restatements, and we use the later one.
What do these two restatment dates correspond to? A restatement is filed with the SEC on one particular date, right?
So between the period end date and the later of the restatement dates, you'll see the older data, and after the later restatement date, you'll see the newer data
Did you mean to write "original filing date" instead of "period end date"? Because the original filing date is usually delayed quite a bit compared to the period end date
pay closer attention post-2019 because the data is better quality / more correct? Or because its got more look ahead bias and can't be trusted?
That's when we switched to Factset. We started snapshotting preliminary data until they flag the filing as complete. So we keep around two versions of the filing. Factset only has one, which gets overwritten when final data is processed. However, our snapshot is not perfect: data can (and does) flow between announcement and filing.
Then again, no point-in-time dataset will ever be perfect. For example S&P has item by item effective dates. But that's the date when they add the item to their databases. For most companies except the main ones, S&P analysts take a few days to process and standardize the filing. For small caps it can be a week or more.
So trying to value of a company immediately after earnings is a dream no matter which vendor you have. Add on top of that the volatility that happens right after, and it's hopeless IMHO. Better off to let it settle down for a week or two.
We're going to add more ways to lag the data soon, which we think is the more correct and prudent approach.
Pretty sure the reason had to do with timing. We could not reliable determine when they happened. We're going over the notes.
2 Likes
Could you also look into what Yuval wrote above, Marco? I don't understand why there would be two restatement dates for a single restated financial statement.
OK so compustat is still the best, most accurate PIT (because the data takes a couple days to be inputted, and they're the only ones to reflect that)?
There are many other financial data providers that have PIT fundamental data. As I understand it, FactSet is one of the only ones that doesn't have PIT fundamentals.
i would support doing a lag as Marco states this will help the data be more PIT like. I can attest to the challenges of finding a true PIT datasource. We are using Visable Alpha data, which S and P just bought, and they are even worse than Factset from a PIT perspective, but much better from a selection of fundamental data.
1 Like
Isn't Visible Alpha for Estimates only? I thought S&P used either CapIQ or Compustat for fundamentals, both of which are PIT.
This is getting into very complex logic details. Only who implemented can explain it.
The problem with Factset restatements is that
- There's only one restatement per filing
- There are only two date fields in the restatement: source date and announce date
We had to make some assumptions:
- Filings are only restated in subsequent filings (true most of the time)
- Filings usually restated once (true most of the time)
- We look for the best match for a subsequent filing that restated the data
- Restated data is only surfaced on the source date or announce date of the subsequent filing, in that order.
Lastly, specially with older dates, sometimes the source date is missing, other times the announce date is missing. And sometimes the dates don't make any sense. So there's additional logic for that.
Hopefully you get the picture. But we're talking about a very small subset of the data. Insignificantly small IMHO.
1 Like
visable alpha has estimates and actuals
1 Like
Yes, I acknowledge that and thank you for taking the time to respond. My use case may be unique: I'm a FactSet Workstation subscriber myself but don't have access to their backtesting software, that's why I'm now a P123 subscriber. But I do need to understand these details for some of my work.
Your explanation of the two restatement dates was very clear, thanks. But I'm still unclear on why interim restatements are ignored?
We have kept track of when we get data from FactSet, so post 2019 the effective date is real and point-in-time. Prior to when we switched to FactSet it's not. Compustat's effective dates are point-in-time.
The reason we switched from Compustat to FactSet is that Compustat insisted that all Portfolio123 users get a license from them, which normally runs more than $25K per year per user.
1 Like
ahhh interesting so you guys might be sitting on one of the best PIT databases out there.
1 Like