diff --git a/crates/wickra-data/src/csv.rs b/crates/wickra-data/src/csv.rs
index c9f5b212..904405b8 100644
--- a/crates/wickra-data/src/csv.rs
+++ b/crates/wickra-data/src/csv.rs
@@ -3,6 +3,11 @@
 //! The reader is generic over the column layout, but ships with a sensible
 //! default ("timestamp,open,high,low,close,volume") that matches the standard
 //! Binance / Yahoo Finance / kaggle dataset format.
+//!
+//! The reader is defensive about real-world files: a leading UTF-8 byte-order
+//! mark is stripped, surrounding whitespace is trimmed from every field, and a
+//! file whose header does not name the required columns is rejected with a
+//! clear [`Error::Malformed`] instead of silently consuming its first data row.
 
 use std::path::Path;
 
@@ -11,6 +16,12 @@ use serde::Deserialize;
 use crate::error::{Error, Result};
 use wickra_core::Candle;
 
+/// Column names the default OHLCV layout requires. The CSV header must contain
+/// every one of these (extra columns are ignored); matching is exact and
+/// case-sensitive because the underlying `serde` deserialization maps header
+/// names to [`DefaultRow`]'s fields by name.
+const REQUIRED_COLUMNS: [&str; 6] = ["timestamp", "open", "high", "low", "close", "volume"];
+
 /// Default OHLCV CSV row layout.
 ///
 /// The timestamp is parsed as an `i64`; if your file ships an RFC3339 / ISO8601
@@ -39,35 +50,157 @@ impl DefaultRow {
     }
 }
 
+/// A [`std::io::Read`] adapter that transparently skips a leading UTF-8
+/// byte-order mark.
+///
+/// Spreadsheet exporters — Excel in particular — prefix CSV files with the
+/// three-byte UTF-8 BOM `EF BB BF`. Left in place it becomes part of the first
+/// header name (`\u{feff}timestamp`), which then fails to match the
+/// `timestamp` column. This adapter drops the BOM before the CSV parser ever
+/// sees it; files without a BOM pass through untouched.
+#[derive(Debug)]
+pub struct BomStripReader<R> {
+    inner: R,
+    /// Whether the leading bytes have been inspected for a BOM yet.
+    checked: bool,
+    /// Bytes read during BOM detection that turned out *not* to be a BOM and
+    /// must still be handed to the consumer.
+    leftover: Vec<u8>,
+    leftover_pos: usize,
+}
+
+impl<R: std::io::Read> BomStripReader<R> {
+    /// Wrap `inner`, stripping a leading UTF-8 BOM on the first read.
+    pub fn new(inner: R) -> Self {
+        Self {
+            inner,
+            checked: false,
+            leftover: Vec::new(),
+            leftover_pos: 0,
+        }
+    }
+
+    /// On the first read, consume up to three bytes and decide whether they
+    /// form a BOM. A BOM is discarded; anything else is buffered for replay.
+    fn check_bom(&mut self) -> std::io::Result<()> {
+        if self.checked {
+            return Ok(());
+        }
+        self.checked = true;
+
+        let mut probe = [0u8; 3];
+        let mut filled = 0;
+        while filled < probe.len() {
+            let n = self.inner.read(&mut probe[filled..])?;
+            if n == 0 {
+                break; // short source — fewer than 3 bytes total
+            }
+            filled += n;
+        }
+
+        if probe[..filled] != [0xEF, 0xBB, 0xBF] {
+            // Not a BOM (or a short file): replay every probed byte verbatim.
+            self.leftover.extend_from_slice(&probe[..filled]);
+        }
+        Ok(())
+    }
+}
+
+impl<R: std::io::Read> std::io::Read for BomStripReader<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        self.check_bom()?;
+        if self.leftover_pos < self.leftover.len() {
+            let n = (self.leftover.len() - self.leftover_pos).min(buf.len());
+            buf[..n].copy_from_slice(&self.leftover[self.leftover_pos..self.leftover_pos + n]);
+            self.leftover_pos += n;
+            return Ok(n);
+        }
+        self.inner.read(buf)
+    }
+}
+
+/// Validate that a CSV reader's header row names every required OHLCV column.
+fn validate_headers<R: std::io::Read>(reader: &mut csv::Reader<R>) -> Result<()> {
+    let headers = reader.headers()?;
+    let present: Vec<String> = headers.iter().map(|h| h.trim().to_string()).collect();
+    let missing: Vec<&str> = REQUIRED_COLUMNS
+        .iter()
+        .copied()
+        .filter(|col| !present.iter().any(|h| h == col))
+        .collect();
+    if !missing.is_empty() {
+        return Err(Error::Malformed(format!(
+            "CSV header is missing required column(s) [{}]; found [{}] — \
+             the first line must be a header naming {}",
+            missing.join(", "),
+            present.join(", "),
+            REQUIRED_COLUMNS.join(",")
+        )));
+    }
+    Ok(())
+}
+
 /// Streaming OHLCV CSV reader.
 #[derive(Debug)]
 pub struct CandleReader<R: std::io::Read> {
     reader: csv::Reader<R>,
 }
 
-impl CandleReader<std::fs::File> {
-    /// Open a CSV file at `path`. The first line is treated as a header by default.
-    pub fn open<P: AsRef<Path>>(path: P) -> Result<Self> {
-        let reader = csv::ReaderBuilder::new()
+impl<R: std::io::Read> CandleReader<R> {
+    /// Build a trimming CSV reader around `inner` and validate its header.
+    fn build(inner: R) -> Result<Self> {
+        let mut reader = csv::ReaderBuilder::new()
             .has_headers(true)
-            .from_path(path)?;
+            .trim(csv::Trim::All)
+            .from_reader(inner);
+        validate_headers(&mut reader)?;
         Ok(Self { reader })
     }
 }
 
-impl<R: std::io::Read> CandleReader<R> {
-    /// Build a reader from any [`std::io::Read`] source.
-    pub fn from_reader(inner: R) -> Self {
-        Self {
-            reader: csv::ReaderBuilder::new()
-                .has_headers(true)
-                .from_reader(inner),
-        }
+impl CandleReader<BomStripReader<std::fs::File>> {
+    /// Open a CSV file at `path`.
+    ///
+    /// The first line must be a header row naming the OHLCV columns; a leading
+    /// UTF-8 BOM and whitespace around values are tolerated.
+    ///
+    /// # Errors
+    /// Returns [`Error::Io`] if the file cannot be opened and
+    /// [`Error::Malformed`] if the header does not contain every required
+    /// column (`timestamp,open,high,low,close,volume`).
+    pub fn open<P: AsRef<Path>>(path: P) -> Result<Self> {
+        let file = std::fs::File::open(path)?;
+        Self::from_reader(file)
     }
+}
 
-    /// Replace the underlying reader; useful for testing.
-    pub fn from_csv_reader(reader: csv::Reader<R>) -> Self {
-        Self { reader }
+impl<R: std::io::Read> CandleReader<BomStripReader<R>> {
+    /// Build a reader from any [`std::io::Read`] source.
+    ///
+    /// A leading UTF-8 BOM is stripped and the header is validated.
+    ///
+    /// # Errors
+    /// Returns [`Error::Malformed`] if the header does not contain every
+    /// required column.
+    pub fn from_reader(inner: R) -> Result<Self> {
+        Self::build(BomStripReader::new(inner))
+    }
+}
+
+impl<R: std::io::Read> CandleReader<R> {
+    /// Wrap a pre-built [`csv::Reader`]; useful for testing or for non-default
+    /// reader configuration.
+    ///
+    /// Unlike [`from_reader`](Self::from_reader) this does *not* strip a BOM —
+    /// the caller owns the reader's configuration — but the header is still
+    /// validated.
+    ///
+    /// # Errors
+    /// Returns [`Error::Malformed`] if the header does not contain every
+    /// required column.
+    pub fn from_csv_reader(mut reader: csv::Reader<R>) -> Result<Self> {
+        validate_headers(&mut reader)?;
+        Ok(Self { reader })
     }
 
     /// Iterator over decoded candles.
@@ -122,8 +255,71 @@ mod tests {
     #[test]
     fn from_reader_works_on_in_memory_data() {
         let data = "timestamp,open,high,low,close,volume\n1,1,2,0,1,10\n2,1,2,0,1,10\n";
-        let mut r = CandleReader::from_reader(data.as_bytes());
+        let mut r = CandleReader::from_reader(data.as_bytes()).unwrap();
         let v = r.read_all().unwrap();
         assert_eq!(v.len(), 2);
     }
+
+    #[test]
+    fn rejects_file_without_header() {
+        // No header row — the first line is data. Without validation the
+        // reader would silently swallow it as the header.
+        let data = "1,10.0,11.0,9.0,10.5,100\n2,10.5,11.5,10.0,11.0,150\n";
+        let err = CandleReader::from_reader(data.as_bytes()).unwrap_err();
+        assert!(matches!(err, Error::Malformed(_)));
+    }
+
+    #[test]
+    fn rejects_header_missing_a_column() {
+        // "volume" is absent.
+        let data = "timestamp,open,high,low,close\n1,10.0,11.0,9.0,10.5\n";
+        let err = CandleReader::from_reader(data.as_bytes()).unwrap_err();
+        match err {
+            Error::Malformed(msg) => assert!(msg.contains("volume"), "msg: {msg}"),
+            other => panic!("expected Malformed, got {other:?}"),
+        }
+    }
+
+    #[test]
+    fn strips_leading_utf8_bom() {
+        // A BOM (\u{feff}) prefixes the header — Excel exports look like this.
+        let data = "\u{feff}timestamp,open,high,low,close,volume\n1,10.0,11.0,9.0,10.5,100\n";
+        let mut r = CandleReader::from_reader(data.as_bytes()).unwrap();
+        let v = r.read_all().unwrap();
+        assert_eq!(v.len(), 1);
+        assert_eq!(v[0].timestamp, 1);
+        assert_eq!(v[0].open, 10.0);
+    }
+
+    #[test]
+    fn tolerates_whitespace_around_fields() {
+        let data = " timestamp , open , high , low , close , volume \n\
+                     1 , 10.0 , 11.0 , 9.0 , 10.5 , 100 \n";
+        let mut r = CandleReader::from_reader(data.as_bytes()).unwrap();
+        let v = r.read_all().unwrap();
+        assert_eq!(v.len(), 1);
+        assert_eq!(v[0].close, 10.5);
+        assert_eq!(v[0].volume, 100.0);
+    }
+
+    #[test]
+    fn bom_stripper_passes_through_non_bom_input() {
+        use std::io::Read;
+        let mut out = String::new();
+        BomStripReader::new("hello".as_bytes())
+            .read_to_string(&mut out)
+            .unwrap();
+        assert_eq!(out, "hello");
+    }
+
+    #[test]
+    fn bom_stripper_handles_short_input() {
+        use std::io::Read;
+        let mut out = Vec::new();
+        // Two bytes — shorter than a 3-byte BOM.
+        BomStripReader::new([0x41u8, 0x42u8].as_slice())
+            .read_to_end(&mut out)
+            .unwrap();
+        assert_eq!(out, vec![0x41, 0x42]);
+    }
 }