For six days straight, 13 to 18 June, the 7 day forecast on the site showed 0% chance of rain. The Dayboro Model was actually rating most of those days well above 50%, up to 77%. It knew rain was likely. The website was reading the wrong number. That is the short version. Here is waht was actually happening.

Two outputs from one run

Every time the Dayboro Model runs, it produces two things. First, a written summary, the prose you can read in the forecast text. Second, a clean daily data table, one row per day, with a proper 24 hour rain probability for every single day in the period. The model always fills in that table. Always. That number exists for every day.

What I had not noticed was that the website was not reading from the data table. It was trying to pull the rain chance out of the prose summary instead. The model writes that prose sentence, something like "chance of precipitation 30 percent", for some forecast periods but not all. On days where the sentence did not appear, the site had nothing to parse and quietly defaulted the rain chance to 0%.

So for those six days, 13 to 18 June, the model was sitting there with its probability estimates of 77%, 66%, 74%, 68%, 57% and 33%, and the website was showing visitors 0% for every one of them.

I control this kind of thing fairly regularly now but I missed this one for too long. Fair enough. It is fixed.

What I changed

I changed the website to read from the clean data table directly. Not the prose. The table is what the model actually produces as output. The prose is a formatted summary of that output, written for humans to read, and it was never meant to be the source of truth for the numbers displayed on screen.

Same model. Same forecast run. Same data. Just the correct number instead of a number I was trying to scrape out of a sentence that sometimes was not there.

Every day in the 7 day forecast now shows the Dayboro Model's actual rain probability for that 24 hour period. Not an approximation, not a default, the real number.

What this means for accuracy tracking

Sicne this fix also gave me access to clean probability numbers for every day going back through the model's history, I added proper reliability tracking. This is sometimes called calibration in the forecasting world. The question it asks is: when the model says 70% chance of rain, does it actually rain about 70% of the time? That is what a reliable probability forecast should do.

The early numbers show the model is well calibrated in the high probability range. When it says 80%, it rains about 90% of the time, which is actually better than 80%. In the 70 to 90% range, I reckon the model is performing well.

But there is a problem at the low end that I am not going to hide. Days the model rated near 0% still rained about 40% of the time. That is a real miss. The model tends to give a very low probability on days that still produce rain. It says "almost certainly dry" when it should not. We are now tracking this systematically, which is a step I could not take until the probability numbers were being read correctly.

I do not have enough data yet to say whether that low end problem is a consistent bias or a pattern tied to specific conditions. That is what the tracking is for. It will take a few months of data to say anything solid.

One more honest thing about the rainfall numbers

The probability, the 60% or 80% chance of rain, is the reliable part of the forecast. That is the number to trust.

The millimetre estimate is a rougher guess. Our model produces an expected rainfall amount for each day and we show it, but the variability around that number is high. In practice, the correct question to ask the forecast is "will it rain?" not "exactly how much will fall?" The probability answers the first question reasonably well. The mm figure answers the second question with less confidence.

I mention this because I think members deserve to know what the model is good at and what it is not. The probability is there now for all 7 days, reading from the right place, and we are tracking whether it earns your trust over time. The mm column is an educated guess.

Why I am telling you this

A member noticed the forecast felt wrong. It had been raining for days and the site was showing 0%. They were right to think something was off.

When I dug into it, the problem was real and simple. The fix was also simple once I found it. I could have quietly pushed the fix and moved on. I would rather tell you what was wrong because it is your forecast and you should know how it works, where it has been wrong, and what we are doing about it.

The tracking is now running. If the Dayboro Model's probability estimates are reliable, the data will show that over the next few months. If they are not, the data will show that too. Either way, I will tell you.

Henk