Topics:
Analyses on last week’s forecasts concerning the meeting between the President of Russia Vladimir Putin and the U.S. President Donald Trump.
The likelihood of a peace in Ukraine.
Forecasts for U.S. inflation.
We start this week’s forecasts by assessing the scenarios we and Grok 4 AI presented a week ago for the meeting between President Trump and Putin. We wanted to test how an untrained AI copes with a task with high uncertainties. Results were mixed, heavily. We wanted to test (an untrained) AI and also to warn you about relying on AI (large language models) in analysis and forecasting work. They are magnificent assistants but untrustworthy masters. After this we turn to providing the analysis and likelihood for peace in Ukraine.
U.S. inflation has picked up, like we warned in May it would. We provide an update to our forecasts using a Vector Autoregression (VAR) framework. This is simply because we had no time to use more sophisticated and better-suited models for this report (we will explain this in detail below). We are moving in that direction, though, and hope to have a larger (more accurate) model for forecasting U.S. inflation by the year-end.
I also want to inform you that we will begin developing an AI-based GDP growth forecasting model with the assistance of an AI researcher. The idea is to train the AI to spot the economic and financial turning points from a swathe of data. This will naturally take some time, considering our limited resources, but we will keep you updated on all (interesting) developments.
Tuomas
The accuracy of the forecasts on the meeting between Presidents Putin and Trump
A week ago, we provided forecasts for the outcome of the meeting between Presidents Trump and Putin occurring later that day. We also asked Grok 4 AI about the outcome. On Wednesday, August 13, Grok 4 saw five scenarios for Friday’s talks (with likelihoods in parentheses):
Scenario 1: Breakthrough Agreement on Ceasefire Framework (likelihood: medium)
Scenario 2: Stalemate with Continued Negotiations (high)
Scenario 3: Escalation and Breakdown (low)
Scenario 4: Surprise Inclusion of Broader Issues and Partial Wins (high)
Scenario 5: Unforeseen Disruption Leading to Postponement (very low)
The explanation for Scenario 1 (the closest what appeared) was:
The meeting concludes with a joint statement outlining a preliminary ceasefire deal, potentially involving territorial "swaps" where Russia retains control over parts of eastern Ukraine in exchange for halting advances and withdrawing from other areas. Trump hails it as a "historic win" on Truth Social, emphasizing his role in ending the "bloodbath," while Putin frames it as a recognition of Russia's security concerns. Follow-up talks involving Zelenskyy are scheduled, but Ukraine expresses reservations about concessions. This outcome boosts Trump's domestic image ahead of midterms but draws criticism from European allies for sidelining Kyiv.
However, Grok assumed that the likelihood of this was medium, while it assessed that Stalemate (#2) and Surprise (#4) would have the highest chances of occurring.1 Our scenarios (finalized on Friday) were:
We are effectively somewhere between #3 and #4. As noted by Tuomas on Monday, “Based on, for example, an article by Reuters, #3 of the above is what Presidents Trump and Putin would have agreed on on Friday, but there would be something that is not known yet, hinting to the existence of a secret deal (#4).” The ‘probability mass’ (60%) of our forecasts was clearly tilted towards these two outcomes.
What we wanted to do last week was to test how accurately an untrained AI (a large language model) is able to forecast outcomes of highly uncertain negotiations. While Grok 4 AI had a clue what was coming, it missed the likelihood of the scenarios quite heavily, and there was no deeper analysis. Its scenarios also followed the common (propagandist) Western narrative. For us, this experiment underlined our previous assessment that, while AI in its current state is helpful in checking facts and theories, deduction capabilities are constrained by the data and commands that are fed to it. For the AI to produce reliable analyses on complex phenomena, you need to be able to program it by yourself. You simply cannot trust most of the analysis AI models produce. Especially in forecasting, modern, untrained AI should be left only to provide assistance.
Let’s turn now to analyzing what the meeting on Friday produced.
The likelihood of a peace in Ukraine
This is how Grok 4 saw the likelihood of a peace on Tuesday, August 19, at 11:30 am Helsinki-time.
There are again some caveats in this forecast, and forecasts of most (human) analysts, for two main reasons:
