How the Sager forecast works

The line under the live readings on the front page doesn't come from a weather model. It comes from a mechanical instrument you read by hand — four dials, a four-character code, and a lookup table. Here's exactly how it reaches its answer.

The Sager Weathercaster · 12–24 hour outlook

Right now at the plot

The idea

The Sager Weathercaster is a forecasting instrument devised by Raymond Sager. It contains no model of the atmosphere and does no arithmetic. It encodes something older: the observation that the combination of wind direction, how that direction has shifted, air pressure and the trend in that pressure predicts the next half-day rather well in temperate latitudes.

You read four dials, combine them into a code such as C231, and look the code up. There are 5,000 possible codes and the table answers essentially all of them. That's the whole method.

Dial 1 · The wind

The first character carries the most information: not just where the wind is blowing from, but which way it has rotated over the last six hours. In the northern hemisphere a wind that turns clockwise usually means a weather system is clearing; one that turns anticlockwise usually means a front is arriving.

Eight compass sectors × three kinds of change = 24 letters, plus Z for calm. The letter I is skipped, so the run is A–Z with a gap.

Plus Z — calm, under 1 mph, whatever the recorded bearing.

Dial 2 · The barometer

Air pressure now, in hectopascals, in one of eight bands. High pressure settles the weather; low pressure stirs it up.

Dial 3 · The barometer's trend

How that pressure has moved over the last three hours. This is the dial that carries the urgency: a rapid fall is the single most reliable warning of rough weather coming.

Dial 4 · The sky

What the sky is actually doing at this moment — the instrument's one check against reality.

Those abbreviations are from METAR, the international format airports use to report conditions. A report such as METAR EGBB 011920Z 32009KT CAVOK 20/07 Q1020 reads: Birmingham, 19:20 UTC, wind from 320° at 9 knots, sky and visibility OK, 20 °C, pressure 1020 hPa.

Putting it together

The four readings are simply written side by side. Here is the code the front page is showing at the moment, taken apart:

Reading the answer

Every one of the 378 possible forecasts is built from three parts in the same order:

The steadiness figure

Under the forecast sits a count and a slim bar — “2,970 of 3,000”, above the words samples agreed. It answers a narrow question: given how precisely we can actually measure the inputs, how often does this same kind of weather come out?

It is shown as a count rather than a percentage on purpose. “99%” reads as a claim about the weather — this forecast is 99% likely — and it is not one, nor could it be: a rock-steady reading of a chart that later proves wrong still scores 100%. A tally of samples cannot be misread that way. At the low end it also asks the more useful question: if 1,110 of 3,000 agreed, what did the other 1,890 say? Usually the runner-up named beside the forecast.

The denominator is the run’s own figure and is not always exactly three thousand. Any draw that fails to produce a forecast at all is discarded rather than counted as a disagreement, so on a bad reading it can say “of 2,987”.

Every reading arrives with a known error attached. The barometer is a modelled figure carrying one decimal, and the tendency is a three-hour difference taken from the same series; on the airport fallback both are coarser, because a METAR reports whole hectopascals. The wind is an average of a bearing that wanders — and below about 5 mph it wanders so far that the vane is barely measuring direction at all — which is usually what a low count here is telling you.

So rather than pretend the reading is exact, the station jitters each input inside its own error three thousand times, re-reads all four dials each time, and counts how often the same kind of weather comes back. The reading is also allowed to drift by its own trend across the twelve hours the forecast covers, because a barometer falling steadily now will be somewhere else by this evening.

It is scored on the kind of weather — fair, unsettled, clouding over, showers, rain, turning fair — and not on the exact sentence. The wind dial rewrites the wind wording, so scoring word-for-word gives a gloomy number on readings where the weather itself was never in doubt: one live reading scored 40% by sentence and 100% by outlook, and 100% was the honest answer. Both said “Fair”.

What it does not tell you. It counts samples rather than stating a confidence, deliberately: this is not the chance the forecast is right. It measures how firmly our instruments pin the reading down, not whether a 1969 cardboard slide chart is correct about tomorrow. A rock-steady reading of a chart that turns out to be wrong still scores 100%. Measuring real accuracy would mean recording every forecast and checking it against what actually happened, which this station does not yet do.

Try the dials yourself

Every one of the 5,000 codes is in this page. Set the four dials and see what the instrument would say.

Where our readings come from

The Almanac reads the dials from two places, and it is worth being straight about which:

• From the plot itself — wind direction now, wind direction six hours ago, temperature, and whether it has rained in the last half hour.
• From a weather model, computed over the plot — the air pressure, its trend, the state of the sky, and whether anything is falling. The allotment station has no barometer, and no instrument that can see cloud.

So the wind half of the forecast is genuinely local, and the pressure and sky half is modelled at this location rather than measured here.

Why a model and not the airport

It used to read Birmingham Airport, 29 km away, and for a while you could choose between the two. The airport is a real observation, which sounds like it ought to win. The trouble is that it describes an aerodrome, not this plot.

Its report is written for pilots. CAVOK — the code you see on a fine day — means “no cloud below 5,000 ft, visibility over 10 km, and nothing significant falling at Birmingham”. A solid deck at 8,000 ft is invisible to it, because nothing in the code can mention that height. On 18 August 2026 it reported CAVOK for three hours while it drizzled here and the model held 100 % cloud throughout.

A model is fed by satellite, which sees the top of everything, everywhere, rather than a laser pointed straight up at one spot on a runway. That is also why phone apps agreed with the model that evening: they read the same kind of field, not a station report tens of kilometres away.

The airport has not been discarded. If the model cannot be reached the forecast falls back to it rather than showing nothing, and the reading beside the forecast always names the station it came from — so “· EGBB” in place of “· Bromsgrove” is the page telling you the model was unavailable. The barometer gauge and its 24-hour history on the dashboard read the same way — the model over the plot, the airport behind it — and the card names whichever answered.

The two pictures

The panel shows two icons: what the sky is doing now, and what the forecast expects over the next 12–24 hours. They come from different places, which is why they can disagree.

The right-hand icon is Sager’s. His table has seven kinds of outlook — fair, turning fair, unsettled, clouding over, showers, rain, snow — and that is all it can ever say. There is no fog, no sleet and no thunder in a 1969 slide chart, so those never appear on that side however wild the day.

The left-hand icon is the model’s own reading of the present, and it is far more specific: all 28 codes it can report are drawn, including fog, freezing rain and thunderstorms. It also separates a shower from steady rain — a shower falls from broken cloud, which is why that icon has sun or moon in it and the steady-rain one does not.

Only the showery pictures change between day and night. Persistent rain falls from a solid deck, so there is no sun to show and nothing for a night version to differ by. Whether it is day or night comes from the model too, not from the clock.

Two places we depart from the original

• Calm is measured by wind speed, not bearing. The original treats a recorded bearing of exactly 0° as “calm”. We take the wind speed instead, so a genuine northerly is never mistaken for still air.
• The pressure trend is fitted, not subtracted — on the airport fallback. A METAR reports pressure only in whole hectopascals, yet this dial has thresholds at 0.7, finer than the readings themselves. Rather than subtract two rounded numbers and discard the rest, we fit a line through every reading in the window, first discarding any that are plainly wrong. The modelled reading needs none of this — it carries a decimal, so its tendency is a plain three-hour difference — but the machinery stays for the days the airport answers.

The seasonal instruments

The dashboard carries four dials that are not readings. Every other gauge on that page shows what an instrument measured; these take the plot’s own temperature, wind and rain and put them through a published method. They are drawn in the same style, so it is worth saying plainly which is which.

Fire index

The Canadian Forest Fire Weather Index (Van Wagner & Pickett, 1987). It carries three moisture stores that dry at different speeds: fine litter over about two thirds of a day, the duff layer over roughly a fortnight, and the deep compact ground over some fifty days. That is why one wet afternoon drops the top figure sharply while the drought code moves only a little — and only if enough fell to reach the deep ground at all. Light rain never gets that far: below 2.8 mm in a day the drought code does not respond, and above it only the excess counts. On 20 August 2026, 5.1 mm took 36 points off a code of 819, which the following two dry days had all but put back.

An earlier attempt used a simpler index reading temperature and humidity at a single instant. It sat in its lowest band on 29 days out of 30 through a drought, which told you nothing. The point of carrying moisture stores is that they remember.

Water deficit

How much more water the weather asked for than fell, carried forward day by day. Evaporation is estimated by FAO-56 Hargreaves, which works from the day’s temperature range and the sun angle for this latitude — the wider the gap between maximum and minimum, the more the day drew out of the ground.

It is a running balance, not a streak. A wet day pays down the debt by however much it actually gave, and no more: on 20 August 2026 the plot’s gauge took 4.6 mm of rain against under 3 mm of evaporation, and the deficit moved by the couple of millimetres it had earned rather than being wiped out. Only a genuinely wet spell — enough rain to catch up with everything evaporated since — clears it back to nothing.

The balance is carried from the previous 1 October. Everything before it last stood at zero is discarded by the floor, and an English winter puts it there many times over — so October is far enough back to measure any summer’s debt from level ground. If a debt does reach the whole way back to October, the figure is shown as “at least”, because it began before the balance was picked up.

What it is not is a soil moisture reading. It measures what the weather asked for, not how far below normal the ground is; real soil can only give up what it holds, and stops once it is empty. The two part company in a long drought, which is why the fire index beside it — which does model the ground — can sit high while a shower nudges this one.

Shooting conditions

Whether the air is worth shooting an airgun in, read from the plot’s own wind, rain and temperature. It is the only instrument here that is not seasonal — it has something to say every day of the year.

Wind is the whole story for a pellet, which is light and slow, so the card works in inches of drift rather than in miles per hour. Drift comes from the standard lag-time model: a pellet is not carried at the wind’s speed, it is pushed by however much longer the wind has had it than a vacuum would. Muzzle velocity is not a free choice either — both calibres are held to the same legal energy, so a .22 leaves much slower than a .177 simply because it is the same energy spread over nearly twice the mass.

That is why the calibre and power switches change the answer rather than the label. At the 12 ft-lb limit a .22 drifts about 17% less than a .177 at every range, which surprises people: it is the slower pellet, but it sheds speed so much more slowly that it is in the air for less extra time. A 6 ft-lb pistol drifts about 40% more than a rifle with the same pellet, for the mirror reason.

The thresholds are a fraction of the target you are trying to stay inside, which is set at one inch — the same for both guns, because the size of what you are aiming at is a property of the target and not of what you are holding. The pistol still reads harder, but only by the amount the ballistics actually justify. Wind is judged straight across the shot, which is the worst case: the plot knows where the wind is from, but not which way you are facing, and the same wind head-on barely moves a pellet.

That inch is the one number here that is a judgement rather than physics — everything else is derived from the pellet and the legal limit. It is also what makes the card worth reading: over twenty thousand simulated days of this station’s wind, a .177 rifle at thirty yards comes out roughly a quarter in each of the four verdicts. Set it at two inches and the worst verdict never appears at all.

Four tests, and the verdict is the worst of them. Wind is how far the pellet is pushed; steadiness is the part of that you cannot hold off for, because it arrives after the shot is called — a steady eight is easier to shoot than three gusting twelve, and every forecast would call the second one the calmer day. Then rain, on the lens and the hands, and temperature, which costs a pre-charged gun muzzle velocity as its reservoir pressure falls, and puts mirage in a scope when it is hot.

A reading that is merely missing is never treated as one that was fine: short of something decisive, the card says Unknown rather than guessing. And a wind reading of zero is reported as under what the cups will turn, not as a proven calm — a cup anemometer has a starting speed below which it simply does not move, and a frozen one reads the same as a still evening.

Sun hours and wind energy

Sun hours count the daylight hours whose solar reading clears a threshold, from the plot’s own sensor. Wind energy is the cube of wind speed through the standard power formula — cube, because doubling the wind carries eight times the energy, which is why a gusty week counts for so much more than a breezy one.

All four are built from whole days, and a day is not finished until it is over. Evaporation needs the day’s true highest and lowest temperature, and the rain total is still climbing while it rains — so today is always provisional, and the figures only settle once midnight has passed. A shower this afternoon shows up properly tomorrow.

Behind that, the numbers are recomputed roughly hourly rather than on every visit, so a reading may be up to about ninety minutes old. That matters far less than the daily settling: a drought code with a fifty-day memory does not care about ninety minutes.

What this means for the numbers. Each is a model with assumptions built in, fed by one station in one garden. Hargreaves does not know your soil; the fire index was calibrated in Canadian forest, not on a Worcestershire allotment. Read them as well-founded estimates that are consistent with themselves over time — the trend from week to week is worth far more than any single figure.

How accurate is it?

Treat it as a well-informed hunch rather than a promise. The readings it works from are modern — the model behind them is fed by satellite — but the reasoning is a 1969 lookup table that sees only four dials at one moment. It has no radar, no sense of what is approaching from the west, and no idea what any forecaster thinks. Wind shift and pressure trend are genuinely powerful signals, which is why the instrument was taken seriously for decades, and it is at its best warning that settled weather is about to break.

It updates every ten minutes, and looks 12 to 24 hours ahead.