Done putting in some limit orders for the night, @gonnarekt or anyone else if you wanna bet more

@parhizj Thank you my friend!

Vibe betting always works! (lets see how well it works this month):

Expecting 2025 to supercede 2024's record of 1.39 for March based on mostly statistical extrapolation for missing forecast data (last 12 days).

Expecting current forecast up until middle of month to track or be slightly above 2024 on average. The second half of March 2024 had an abnormal dip, so on average following the trend I expect something more like this for March (ignore rest of year past March), using the OU as a guess for the offset:

You can imagine the line being within (very roughly) +- 0.6 C for the last ~2 weeks and roughly +- 0.3 for the next 2 weeks.

Even with a slightly less likely 0 offset for the error for the rest of the month (less likely since I expect it will take some time for the error to drift back to 0, based on the recent trend, giving on average a positive offset for the rest of the month), on average it looks like 2025 March will beat 2024 just by eyeballing it, so at the moment I think the last two bins should be the most likely:

Past March 24 I fill in statistically (you can ignore the rest of the year past March):

The end of the 3rd week for the GEFS forecast shows the temperature rise levelling off, so I look at some statistics for the last week of March in ERA5 given the GEFS forecast now covers until March 24.

Now the statistics for last week of March (the last datapoint in the two charts below is the 0 offset prediction for 2025):

This shows almost always (81/85 = ~95% of the time) the average temperature in the last 7 days of March shows a rise compared to the March 24 temperature (an average of ~0.147 C higher). (The 2025 guess has it about 0.117 higher than March 24 so not too far from the mean and actually a bit under)

With the absolute temp average for last week the last week of March 2025 a bit higher than 2024 (based on extending the GEFS forecast statistically).

As I still need to offset the error for part of the GEFS forecast (although it has been decreasing) I do some basic calculations for a scenario where we assume the offset of 0 is correct for the GEFS portion but the statistical forecast rise is wrong for the last 7 days, and for instance the average temperature from March 25-31 is the same as March 24 or less: this is the 5% chance mentioned above that corresponds to any offset ~ <-0.037 (= -(.117) * (7/22)) in the chart below:

The OU meta-prediction is still predicting a slightly positive offset for the remainder of the month (0-0.09) looks plausible), but at the moment it seems likely (with low confidence) the anomaly for March will still be broken (currently at 1.39 for March 2024) given the GEFS forecast and the above...

Its still too early to say what offset is correct, but the odds are tilted towards still being positive ~0.05:

March is the hardest month to predict though so we shall see...

Polymarket for reference (has completely different bins) and a completely different forecast at the moment (currently predicting something below or near last year's anomaly temp):

@parhizj gopfan saw all of this and went against everything you just said 😂 that dude bats a thousand. gopfan more like goatfan

@LeonardoParaiso That’s the whole point of markets. At least the individual bin probabilities do not seem wildly overconfident to me at the moment; Poly is at 61%.

Did some more work today including deterministic ECM data to try to come up with an even yet more elaborate prediction system combining ECM and GEFS after all the final adjustments are all done by splitting the difference of the average for the remaining forecasted temps.

In doing so I did understand the source of some of the discrepancy at least between what I have been predicting using GEFS for March and what seems clear others are predicting using I believe ECM data:

Comparing the last 5 runs using ECM (wherein I try fit to the ERA5 data and subsequently further adjust the forecast offsets using ARIMA (as using OU did not work out with ECM)) with the GEFS reveals ECM predicts lower temps for the middle of the month finally matching GEFS towards the end of the month.

Using a (mixed) best guess for the offset for GEFS (red line) it looks like this:

I thereafter calculate the average difference for GEFS (with offsets) - each ECM model run (last 5 model runs) across common dates (after all offsets already):

   model_init  avg_diff
0  2025031012  0.153844
1  2025031100  0.203933
2  2025031112  0.152150
3  2025031200  0.110761
4  2025031212  0.175303
Mean of average differences: 0.1592

I then split the difference in offsets to calculate a new split offset (this I hope has the effect of weighting very roughly equal the final adjusted ECM and GEFS (mapped) forecasts)

This corresponds to an offset that is now slightly negative (which more closely matches the last couple errors for gefs predictions - ERA5), which looks like this now:

This corresponds to the grey line as the split difference offset below, which shows roughly equal odds for the highest 3 bins (with the highest bin with more probability mass).

I don't particularly like the look of the last couple few days though from GEFS:

This didn't have any serious consequences for my recent betting in hindsight, but it does show that it is especially sensitive to the next week or so rather than the end of the month (will it continue near-plateuing like ECM, jump up as high as GEFS suggests, or somewhere in between)

For any interest bettors, a simple linear squares regression to debias from ECM->ERA5 leaves quite a bit to be desired (this is only using the 00Z init forecast and 0,6,12,18h valid time samples)

Mean Squared Error (MSE): 0.06
Root Mean Squared Error (RMSE): 0.25
Mean Absolute Error (MAE): 0.21

OU did not validate well but the time series dependencies of the runs on the residuals also is obvious even without running autocorrelation and partial autocorrelation, thus resorting to ARIMA which seems ok but I haven't put any instrumentation to monitor its performance.