TL;DR: Walk-forward validation rewinds the clock to a past date, re-runs a forecast using only the data visible at that moment, and checks it against the figures that were actually reported later — it answers whether an estimation method has been accurate in the past.
Concepts
Why validate at all?
Every forward valuation rests on one estimated number: roughly how much this company will earn this year. The problem is that an estimate does not come with an accuracy label attached.
A beautifully drawn forward price chart may be sitting on an earnings forecast that has overshot by thirty percent for years. The reader cannot tell, because the chart does not say so.
Walk-forward validation answers exactly that question: how accurate has this method been, for this specific company, over the past few years?
What "only the data visible at the time" means
This is the most important part of walk-forward validation, and the easiest to get wrong.
Say you want to validate the 2023 estimate. The correct procedure is to rewind to mid-2023, feed in only the monthly revenue and financial statements published by then, produce a full-year 2023 EPS estimate, and compare it against the actual figure that only appeared in the 2024 annual report.
The wrong procedure is to "estimate" 2023 using data that already contains the 2023 outcome. That is not forecasting, it is copying the answer — the technical name is look-ahead bias, the most common way a backtest cheats, and it makes a bad method look uncanny.
How this differs from a strategy backtest
A strategy backtest asks whether buying and selling this way makes money. Walk-forward validation asks something one layer earlier: is the input number itself trustworthy?
The two can be stacked, but not reordered. If the input is unreliable, any conclusion built on top of it has no foundation to discuss.
In practice: reading it on CTSstock
At the bottom of Stock Analysis › VIP › Scenario › P/E there is a walk-forward validation table. Each row is one fiscal year:
| Column | Meaning |
|---|---|
| Year | The year being replayed |
| Projected | Full-year EPS computed from data available at that time only |
| Actual | The EPS the statements later reported for that year |
| Error | (Projected − Actual) ÷ Actual, as a percentage |
How to read the table
The first thing to look at is not the size of the error but its stability.
- Errors oscillating around plus or minus ten percent — the method is usable for this company; keep a ten percent buffer in mind when reading forward prices.
- Errors jumping around (+3% one year, −40% the next) — this company's earnings are inherently hard to forecast, so treat forward prices as a rough reference only.
- Errors consistently leaning one way (several years of overestimation) — that is a systematic bias. It matters more than a large error, because it is predictable and you can correct for it yourself.
Why only Taiwan-listed stocks have this section
Replaying a forecast requires that "what was visible at the time" can be reconstructed. Taiwan-listed companies publish the previous month's revenue by the tenth of each month, so it is possible to reconstruct exactly what was known in June 2023.
US, Japanese, Korean and Canadian coverage uses trailing four-quarter EPS, and Hong Kong uses annual reports. These markets have no intra-year data at the same frequency to rewind through, so a replay forced onto them would only manufacture look-ahead bias. Better to omit it than to publish a fake one.
FAQ
Q: How small does the error need to be before I trust it?
There is no universal threshold, because it depends on the industry. Cyclical businesses (shipping, panels, memory) swing violently by nature, so a thirty percent error is not outrageous. For utilities and telecoms, an error beyond ten percent suggests something is wrong with the method. Comparing against this company's own history is more meaningful than comparing across companies.
Q: If it was accurate before, will it be accurate ahead?
No. Walk-forward validation only describes what already happened. Cyclical turning points, one-off non-operating items, acquisitions and sharp currency moves can all break a previously reliable method. Its purpose is to calibrate your confidence in the number, not to guarantee it.
Q: Are two or three replayed years enough?
Not really, though still useful. Two or three years can reveal an obvious systematic bias but not stability. The more years covered — ideally spanning both halves of at least one business cycle — the sturdier the conclusion. With a small sample, treat it as a warning light rather than a health report.
Q: If it always overestimates, can I just discount it myself?
Yes, and that is one of the table's intended uses. Two cautions, though: the bias may come from one particular period (the pandemic years, say) rather than being a lasting trait; and discounting only adjusts the earnings figure — the valuation multiple band itself does not move with it.