← Back to dashboard

How to read the dashboard

What the model predicts

The model does not claim to name tomorrow's exact exchange rate — for a freely floating currency that is not possible, and anyone who promises a single precise number is guessing. Instead it forecasts a range: where the USD/PLN rate is likely to be, day by day into the future.

That range is built from measured behaviour of the rate, not from opinions about it:

The range is drawn as a fan of three nested bands. The darkest core is where the rate should be half the time (50%), the middle band four times out of five (80%), and the palest outer band nine times out of ten (90%). The line through the middle is the median. It starts at the latest fix and tilts with the model's balance of risks — its directional opinions from the news and on scheduled events — and the fan moves with it. Read that tilt as the model's lean, not as a tested call: we scored both opinions against realised NBP fixings, and the news-driven lean was right 42% of the time over 73 independent observations, the event views 14 times out of 28. Neither beats a coin. The width of the fan is what we have tested and stand behind.

Chart lines

Where the points sit within a day

Each point on every line is placed at the time of day the forecast cycle finished, not at midnight. The completion time is rounded to the nearest half hour (e.g. a run that finished at 13:04 lands at 13:00, one at 13:28 lands at 13:30). So the whole picture sits at "when the forecast was produced" for each day, and the date labels along the bottom are centered on midday of the day they describe.

Vertical lines

Event markers

Hover a marker for the event name, the model's view in grosze with its confidence, and the source. The colour is the model's opinion, not our forecast. Scored on their own, these views matched the next NBP fixing on 14 of 28 event days — a coin — so we show them with that record next to them. A view tilts the line by impact × confidence — a +2 gr view at 35% confidence moves it by about 0.7 gr — but never widens or narrows the range. What sets the range's width is that a key release is due, not which way the model leans on it.

Trend strips

At the bottom of the chart, horizontal strips mark each future week. Most weeks read grey — flat — and that is the honest outcome rather than a missing feature. A strip is only coloured when the week's balance-of-risks lean clears a signal-to-noise threshold, and on a currency that behaves like a random walk at these horizons it rarely does. Green (upward) and red (downward) mark the weeks where it does: the same tilt the line shows — a lean, not a tested call. Hovering a grey week lists the rate decisions, CPI prints and payrolls scheduled in it — the days the fan widens.

Weekday markers

A thin strip along the bottom of the chart labels each working day (Mon–Fri) across the whole timeline, past and future, centered on midday. Weekends are left blank. Past days carry a faint frame to match the look of the future trend strips; only the future strips are interactive (hover for trend details) — the past weekday frames are purely a visual guide.


Chart controls

Accuracy metric (MAE)

The line above the chart scores the model's past forecasts at the horizon you pick (1, 7 or 14 days ahead):


News section

Below the chart are the 10 stories that weighed most in the model's reading of the day. They explain what was in the news; taken together they also tilt the line a little (see below).

News rating number

Each news tile shows a number in the top-right corner. This is the weight — how strongly this story counted in our reading of the day, on a relative scale. It combines the size of the move our scorer thought the story implied, how confident it was, and how authoritative the source is.

The weight is not how far the forecast line moved. The line does lean with the news, but through an aggregate: stories are grouped by theme, only the strongest read in each theme counts, running conflicts count only above their usual level, and everything fades with a five-day half-life. So the number tells you what the day's news looked like to the model and how the tiles are ranked; the line's actual tilt from all of it together is usually a fraction of the top weight. The width of the fan does not respond to the news at all.

Treat the colour as the reading, not as a prediction. We publish these signs because they are what the model saw; we measured whether they anticipate the next fixing, and they do not (42% over 73 independent observations) — which is why the chart captions the line's tilt as a lean rather than a call.

News tiles are ranked by absolute weight: the first tile is the single most influential piece of news, regardless of direction. Click any tile to read the full article.

Disclaimer

This is a personal test project, not a financial service. It is built entirely from publicly available news sources — public news feeds, agency wires and open economic calendars — processed by an automated model whose output is experimental.

Accuracy is not guaranteed. Every past forecast stays on the chart with the range it was published with, so you can see where the actual rate left it — the track record is shown as it is, misses included, and the header counts how often the rate landed inside.

Nothing here is investment advice or a recommendation. Do not make financial decisions based on this site. The author accepts no liability for any loss arising from the use of this information.