After converting a running file into a table, some cells are empty, others contain zero, and an unusually large number appears in the heart-rate column. A quick automatic fill may tidy the table while changing what its data mean. A missing value differs from a genuinely measured zero, and an invalid code is not necessarily an extraordinary sporting performance.
Tiny cell rules can have major consequences when processing a workout CSV: a different average, concealed dropout, incorrect units or shifted columns. Our short example series explore the key distinctions. The aim is a checking sequence that reveals, before summarising, whether the programme mistakes missing data for movement or an encoded marker for a measurement.
Missingness is information, rather than an inconvenient gap
If no usable sensor data exist at a particular time, we do not know the value with certainty. An empty cell may indicate that lack of knowledge. Filling it with zero may instead claim a known value of precisely zero. These are different statements. A tidy table that looks complete is not necessarily more accurate than one that visibly preserves gaps.
The official FIT protocol description uses invalid values defined for each type. Processing must associate these with the appropriate type and field. Not every number in a file can automatically be treated as a valid measurement. Decoding matters precisely so value and validity can be separated before everyday analysis.
In our heart-rate log, the sensor briefly loses connection. Correctly marked missing data do not mean the runner’s heart stopped beating. A technical dropout should not be displayed as a physiological zero. A graph gap is then more honest than a line plunging to zero and returning, suggesting that the programme actually measured that change.
The same four timestamps can produce very different averages
In our equally spaced four-point example, three valid values are 140, 145 and 150; the fourth is missing. The simple mean of the three known values is 145. Replacing the missing value with zero gives a four-number mean of 108.75. One artificial fill creates a difference of more than thirty units. No new measurement occurred, yet the summary changed substantially.
If we fill the missing point with the preceding value, 145, the example’s mean is again 145. A convenient result does not prove that the missing value was actually 145. Filling introduced an assumption. Processing must therefore distinguish original and imputed data even when the average apparently stays unchanged. Information’s origin should not disappear without trace into the final number.
Leaving out missing points does not automatically resolve every problem either. If more data are missing from a harder section, what remains may not represent the whole workout. First establish where and how much information is available. “145 is the mean of valid points” is more accurate than “the whole run’s heart rate was definitely 145 on average” when coverage is unknown.
An overly raw export may show an invalid code
In the FIT type table, for example, the uint8 invalid code is 0xFF, or decimal 255. This is a general type definition, rather than a magic number for every column and file format. Other types may use other markers. “Always delete 255” can therefore be as wrong as treating every 255 as a real measurement.
In our technical example, a decoder writes the invalid uint8 code unchanged as 255 in a column. A spreadsheet cannot automatically know not to average it. Ignoring the field’s type turns the marker into a striking spike. The error is lost decoding meaning, rather than a sudden change in the runner.
Some types reserve zero as an invalid value too. That does not mean every numerical zero is missing. The specific format and field definition decide. Processing must therefore be correct field by field, rather than run one deletion condition across the whole table. Identical-looking cells can originate from different data types.
Enlarge imageGenuine zero may also need preserving
Zero can be valid in some fields. Speed during a moment without movement differs from a timestamp with no speed data. Likewise, the meaning of zero in a supported power field depends on the particular device and field description. Do not delete every zero because it indicated dropout in another column.
Our fictional speed series is 3, 0 and 3 metres per second across three equal-length windows. If the middle zero genuinely means standing still, the simple whole-time average is two metres per second. Excluding zero gives three, which characterises only the two moving windows. Both figures may be interpretable, but answer different questions. State zero filtering in the average’s name too.
A switched-off sensor, however, does not prove zero power or movement. If the recording only shows a sensor disappearing, treat it as a data state. The runner may have continued moving while one measure failed to arrive. Other fields can help explain the situation, but cannot automatically provide a certain value for the missing measurement simply to complete a column.
A comma can be both decimal mark and field delimiter
The RFC 4180 description of the common comma-separated CSV format also explains quoting fields that contain commas. In practice, exporters may use another delimiter, such as a semicolon. Identify the column delimiter separately from the decimal mark when opening a file. The extension alone does not provide complete import settings.
In our example, speed is written as 3,5 and heart rate as 145. If the exporter uses decimal commas and semicolon delimiters, the row can contain two fields. If the receiver mistakenly treats every comma as a column boundary, 3,5 splits apart. Following data may enter the wrong columns. The table can still look rectangular, especially if the programme fills additional cells automatically.
Check a few known sample rows for a correct import. Are speed, heart rate and timestamp under their proper headers, and imported as numbers or text? Do not randomly “correct” 3,5 to 35 by removing the comma. Choose locale and delimiter settings appropriate to the original format. Bulk find-and-replace can easily create more errors than it fixes.
Without units, even the value’s scale can be misunderstood
The Garmin guide to custom developer fields separately addresses the field name, data type and unit description. These meanings belong with the number. Preserving only numerical columns when exporting standard and custom fields is insufficient if units are later lost. A graph axis means something only when we know the quantity it represents.
In our example, speed is 3.5 metres per second. That is 12.6 kilometres per hour because an hour contains 3,600 seconds and a kilometre one thousand metres. Both numbers describe the same movement. Incorrectly labelling 3.5 as kilometres per hour implies a very different pace. Correct units then matter more than adding decimal places.
Raw encoded values and already scaled numbers are different states too. A decoder may perform the format’s conversion, while another view shows storage values. Scaling an already converted number again is as wrong as omitting a required conversion. Export documentation and a few checked fields help establish which state you received.
Row count is not automatically the number of seconds
The FIT decoding guide separately discusses uneven recording, records and timer events. Two consecutive CSV rows are not necessarily one second apart. With unequal intervals, a simple row-based average may answer a different question from a time-weighted summary. Timestamps are part of the data alongside valid cells.
In our sample, one value is considered valid for ten seconds and another for only one second under an explicitly stated holding assumption. A simple two-row mean gives equal weight; a time-weighted mean does not. The difference comes from unequal durations associated with rows, rather than an empty cell. The processing goal and signal interpretation determine the appropriate method.
Blindly filling a section missing because of a pause can make the recording look like continuous movement. Conversely, treating every gap between sparse points in continuous movement as a pause removes time from analysis. Timer events and recording mode can help distinguish these cases. Continuous row numbering alone cannot resolve them.
Enlarge imageA five-minute check can prevent a wrong monthly summary
Choose a short, familiar recording section with continuous movement and an identifiable stop or dropout. Inspect the same timestamps in the original app and export. Check headers, units, delimiters and missing-value notation. You need not manually review the whole month to reveal an error in a general import rule.
Our checklist has four separate questions: which cell is a valid measurement, which is a genuine zero, which is missing and which was imputed later? Each can have its own marker in the working table. Preserve the original column and name cleaned and derived columns separately. That keeps the reason an average changed traceable later.
Briefly document the processing rule too. For example: invalid codes were treated as missing according to the field definition, genuine zeros were retained, missing values were not filled, and results were described as applying to valid sections. This is an analysis example, rather than a universal recipe for every data type. The point is a traceable decision appropriate to the field.
A usable log does not require missingness to disappear
Some workout fields may be complete and others partial. Distance and time may remain assessable when part of heart rate is missing. You need not discard the whole run, but conclusions relying on the incomplete column are limited by the data. A good log makes that distinction visible. Automatically filled rows that look complete can more easily create false confidence.
If the exporter’s behaviour is undocumented, note the uncertainty and preserve the unchanged original. Another appropriate processor may later offer another view if original data remain. Manual table corrections must not retrospectively become an original measurement history. Keep correction as a separate, checkable working copy.
The distinction between zero and missing data is therefore part of what workout information means. Checking invalid codes, decimal marks, units and timestamps together helps prevent misleading averages. The aim is data whose claims we understand, rather than a perfectly filled table. Only then should the numbers be used to interpret a runner’s performance.
Sources
- Garmin FIT protocol: data types and invalid values.
- Garmin FIT decoding: timestamps, uneven recording and CSV.
- Garmin FIT Developer Data: field name, data type and units.
- RFC 4180: CSV format and quotation.
How Futószív articles are prepared: our editorial guidelines.


